# zoo-bench performance report

raw data: `results/zoo-framework-0.11.0/20261011T093730Z.json`

## Run environment

| item | value |
|---|---|
| CPU model | AMD EPYC 7763 64-Core Processor |
| logical cores | 4 |
| platform | Linux-x86_64 |
| operating system | Linux 6.17.0-1022-azure (#22-Ubuntu SMP Mon Jul 27 17:24:03 UTC 2026) |
| Python | CPython 3.13.16 |
| interpreter | /opt/hostedtoolcache/Python/3.13.16/x64/bin/python |
| GIL | default |
| framework under test (distribution metadata) | 0.11.0 |
| framework under test (module __version__) | 0.11.0 (footnote only, not the version criterion) |
| drive-surface generation of the framework under test | current-generation drive surface (scheduling core + scheduling model; the completion signal comes from the result reactor) |
| harness | zoo-bench 0.0.1 @ 9a42be24ac43f75e2e5d7b16434f91ba3d0717b6 |
| command to reproduce | `zoo-bench run --framework zoo-framework==0.11.0` |

> dist_version (the distribution metadata) is the source of truth; module_version is only a footnote

## Workload

- composition: one unit of work = JSON serialize + deserialize + string concatenation, repeated until the target duration is reached; the tiers are declared by body_tiers_us in matrix.yaml
- **the workload is a stand-in, not a real trace**: its shape resembles business actions such as 'take a frame of data -> serialize -> parse -> transform strings', but this repository holds no real trace, so every conclusion is conditional on that

## Measurement methodology

- **the body duration is instrumented within the same run**: the body measures itself internally and reports back; framework overhead = end-to-end - it. **No cross-run subtraction**  -  two runs differ in state (caches, frequencies, scheduling noise), so subtracting introduces systematic bias.
- **warmups are excluded from statistics**: each unit first runs 3 warmup rounds that are discarded, then only the following 30 measured rounds are counted; the sample count equals 'measured rounds x concurrency'.
- **absolute durations are not comparable across runs**: they only show the order of magnitude; across versions/machines read the **relative quantities within the same run** (overhead share, multiples against the baselines), which are insensitive to overall speed.

## Conclusion summary

- Once the framework under test zoo runs a body of about 2700 microseconds or more, its framework overhead drops below 15% of end-to-end; below that tier, the main price of choosing it is the framework overhead itself.
- At concurrency 1 and a 40 microsecond body, the framework under test is 2.14x faster than process_pool (the fastest measured tier at that concurrency).
- At concurrency 1, once the body exceeds about 40 microseconds, apscheduler is no longer faster than the framework under test.
- At concurrency 4, once the body exceeds about 40 microseconds, apscheduler is no longer faster than the framework under test.
- At concurrency 16, once the body exceeds about 40 microseconds, apscheduler is no longer faster than the framework under test.
- At concurrency 64, once the body exceeds about 40 microseconds, apscheduler is no longer faster than the framework under test.
- At concurrency 1, within the measured tiers asyncio_pool is **always faster** than the framework under test  -  tiers like these are exactly the openly published unfavourable data of this report.
- At concurrency 4, once the body exceeds about 10000 microseconds, asyncio_pool is no longer faster than the framework under test.
- At concurrency 16, within the measured tiers asyncio_pool is **always faster** than the framework under test  -  tiers like these are exactly the openly published unfavourable data of this report.
- At concurrency 64, within the measured tiers asyncio_pool is **always faster** than the framework under test  -  tiers like these are exactly the openly published unfavourable data of this report.
- At concurrency 1, within the measured tiers bare_thread is **always faster** than the framework under test  -  tiers like these are exactly the openly published unfavourable data of this report.
- At concurrency 4, within the measured tiers bare_thread is **always faster** than the framework under test  -  tiers like these are exactly the openly published unfavourable data of this report.
- At concurrency 16, once the body exceeds about 40 microseconds, bare_thread is no longer faster than the framework under test.
- At concurrency 64, within the measured tiers bare_thread is **always faster** than the framework under test  -  tiers like these are exactly the openly published unfavourable data of this report.
- At concurrency 1, once the body exceeds about 40 microseconds, process_pool is no longer faster than the framework under test.
- At concurrency 4, within the measured tiers process_pool is **always faster** than the framework under test  -  tiers like these are exactly the openly published unfavourable data of this report.
- At concurrency 16, within the measured tiers process_pool is **always faster** than the framework under test  -  tiers like these are exactly the openly published unfavourable data of this report.
- At concurrency 64, once the body exceeds about 40 microseconds, process_pool is no longer faster than the framework under test.
- At concurrency 1, within the measured tiers thread_pool is **always faster** than the framework under test  -  tiers like these are exactly the openly published unfavourable data of this report.
- At concurrency 4, within the measured tiers thread_pool is **always faster** than the framework under test  -  tiers like these are exactly the openly published unfavourable data of this report.
- At concurrency 16, within the measured tiers thread_pool is **always faster** than the framework under test  -  tiers like these are exactly the openly published unfavourable data of this report.
- At concurrency 64, within the measured tiers thread_pool is **always faster** than the framework under test  -  tiers like these are exactly the openly published unfavourable data of this report.

> a single-point speedup has no selection meaning; the conclusion is phrased as 'which scheme to choose at what body duration' (the same framework overhead is 70% of a 40 microsecond task and a few percent of a 10 millisecond one)

### Overhead threshold crossings

| scheme | concurrency | threshold | first tier at or below the threshold (microseconds) | measured tiers (microseconds) | note |
|---|---|---|---|---|---|
| apscheduler | 1 | 15.00% | 2700 | 40, 300, 2700, 10000 | this is the smallest measured tier that satisfies the condition; the true crossing point lies between it and the previous tier |
| apscheduler | 4 | 15.00% | - | 40, 300, 2700, 10000 | this concurrency exceeds the parallel capacity of the machine this run was on: the body's self-reported duration contains scheduling waits from oversubscription, which is not the same thing as 'wall clock / concurrency', so subtracting them is meaningless (measured, it produced negative values). End-to-end, throughput and the body's self-reported values remain valid, so they are given as usual. |
| apscheduler | 16 | 15.00% | - | 40, 300, 2700, 10000 | this concurrency exceeds the parallel capacity of the machine this run was on: the body's self-reported duration contains scheduling waits from oversubscription, which is not the same thing as 'wall clock / concurrency', so subtracting them is meaningless (measured, it produced negative values). End-to-end, throughput and the body's self-reported values remain valid, so they are given as usual. |
| apscheduler | 64 | 15.00% | - | 40, 300, 2700, 10000 | this concurrency exceeds the parallel capacity of the machine this run was on: the body's self-reported duration contains scheduling waits from oversubscription, which is not the same thing as 'wall clock / concurrency', so subtracting them is meaningless (measured, it produced negative values). End-to-end, throughput and the body's self-reported values remain valid, so they are given as usual. |
| asyncio_pool | 1 | 15.00% | 2700 | 40, 300, 2700, 10000 | this is the smallest measured tier that satisfies the condition; the true crossing point lies between it and the previous tier |
| asyncio_pool | 4 | 15.00% | - | 40, 300, 2700, 10000 | this concurrency exceeds the parallel capacity of the machine this run was on: the body's self-reported duration contains scheduling waits from oversubscription, which is not the same thing as 'wall clock / concurrency', so subtracting them is meaningless (measured, it produced negative values). End-to-end, throughput and the body's self-reported values remain valid, so they are given as usual. |
| asyncio_pool | 16 | 15.00% | - | 40, 300, 2700, 10000 | this concurrency exceeds the parallel capacity of the machine this run was on: the body's self-reported duration contains scheduling waits from oversubscription, which is not the same thing as 'wall clock / concurrency', so subtracting them is meaningless (measured, it produced negative values). End-to-end, throughput and the body's self-reported values remain valid, so they are given as usual. |
| asyncio_pool | 64 | 15.00% | - | 40, 300, 2700, 10000 | this concurrency exceeds the parallel capacity of the machine this run was on: the body's self-reported duration contains scheduling waits from oversubscription, which is not the same thing as 'wall clock / concurrency', so subtracting them is meaningless (measured, it produced negative values). End-to-end, throughput and the body's self-reported values remain valid, so they are given as usual. |
| bare_thread | 1 | 15.00% | 2700 | 40, 300, 2700, 10000 | this is the smallest measured tier that satisfies the condition; the true crossing point lies between it and the previous tier |
| bare_thread | 4 | 15.00% | - | 40, 300, 2700, 10000 | this concurrency exceeds the parallel capacity of the machine this run was on: the body's self-reported duration contains scheduling waits from oversubscription, which is not the same thing as 'wall clock / concurrency', so subtracting them is meaningless (measured, it produced negative values). End-to-end, throughput and the body's self-reported values remain valid, so they are given as usual. |
| bare_thread | 16 | 15.00% | - | 40, 300, 2700, 10000 | this concurrency exceeds the parallel capacity of the machine this run was on: the body's self-reported duration contains scheduling waits from oversubscription, which is not the same thing as 'wall clock / concurrency', so subtracting them is meaningless (measured, it produced negative values). End-to-end, throughput and the body's self-reported values remain valid, so they are given as usual. |
| bare_thread | 64 | 15.00% | - | 40, 300, 2700, 10000 | this concurrency exceeds the parallel capacity of the machine this run was on: the body's self-reported duration contains scheduling waits from oversubscription, which is not the same thing as 'wall clock / concurrency', so subtracting them is meaningless (measured, it produced negative values). End-to-end, throughput and the body's self-reported values remain valid, so they are given as usual. |
| process_pool | 1 | 15.00% | 2700 | 40, 300, 2700, 10000 | this is the smallest measured tier that satisfies the condition; the true crossing point lies between it and the previous tier |
| process_pool | 4 | 15.00% | - | 40, 300, 2700, 10000 | this concurrency exceeds the parallel capacity of the machine this run was on: the body's self-reported duration contains scheduling waits from oversubscription, which is not the same thing as 'wall clock / concurrency', so subtracting them is meaningless (measured, it produced negative values). End-to-end, throughput and the body's self-reported values remain valid, so they are given as usual. |
| process_pool | 16 | 15.00% | - | 40, 300, 2700, 10000 | this concurrency exceeds the parallel capacity of the machine this run was on: the body's self-reported duration contains scheduling waits from oversubscription, which is not the same thing as 'wall clock / concurrency', so subtracting them is meaningless (measured, it produced negative values). End-to-end, throughput and the body's self-reported values remain valid, so they are given as usual. |
| process_pool | 64 | 15.00% | - | 40, 300, 2700, 10000 | this concurrency exceeds the parallel capacity of the machine this run was on: the body's self-reported duration contains scheduling waits from oversubscription, which is not the same thing as 'wall clock / concurrency', so subtracting them is meaningless (measured, it produced negative values). End-to-end, throughput and the body's self-reported values remain valid, so they are given as usual. |
| thread_pool | 1 | 15.00% | 300 | 40, 300, 2700, 10000 | this is the smallest measured tier that satisfies the condition; the true crossing point lies between it and the previous tier |
| thread_pool | 4 | 15.00% | - | 40, 300, 2700, 10000 | this concurrency exceeds the parallel capacity of the machine this run was on: the body's self-reported duration contains scheduling waits from oversubscription, which is not the same thing as 'wall clock / concurrency', so subtracting them is meaningless (measured, it produced negative values). End-to-end, throughput and the body's self-reported values remain valid, so they are given as usual. |
| thread_pool | 16 | 15.00% | - | 40, 300, 2700, 10000 | this concurrency exceeds the parallel capacity of the machine this run was on: the body's self-reported duration contains scheduling waits from oversubscription, which is not the same thing as 'wall clock / concurrency', so subtracting them is meaningless (measured, it produced negative values). End-to-end, throughput and the body's self-reported values remain valid, so they are given as usual. |
| thread_pool | 64 | 15.00% | - | 40, 300, 2700, 10000 | this concurrency exceeds the parallel capacity of the machine this run was on: the body's self-reported duration contains scheduling waits from oversubscription, which is not the same thing as 'wall clock / concurrency', so subtracting them is meaningless (measured, it produced negative values). End-to-end, throughput and the body's self-reported values remain valid, so they are given as usual. |
| zoo | 1 | 15.00% | 2700 | 40, 300, 2700, 10000 | this is the smallest measured tier that satisfies the condition; the true crossing point lies between it and the previous tier |
| zoo | 4 | 15.00% | - | 40, 300, 2700, 10000 | this concurrency exceeds the parallel capacity of the machine this run was on: the body's self-reported duration contains scheduling waits from oversubscription, which is not the same thing as 'wall clock / concurrency', so subtracting them is meaningless (measured, it produced negative values). End-to-end, throughput and the body's self-reported values remain valid, so they are given as usual. |
| zoo | 16 | 15.00% | - | 40, 300, 2700, 10000 | this concurrency exceeds the parallel capacity of the machine this run was on: the body's self-reported duration contains scheduling waits from oversubscription, which is not the same thing as 'wall clock / concurrency', so subtracting them is meaningless (measured, it produced negative values). End-to-end, throughput and the body's self-reported values remain valid, so they are given as usual. |
| zoo | 64 | 15.00% | - | 40, 300, 2700, 10000 | this concurrency exceeds the parallel capacity of the machine this run was on: the body's self-reported duration contains scheduling waits from oversubscription, which is not the same thing as 'wall clock / concurrency', so subtracting them is meaningless (measured, it produced negative values). End-to-end, throughput and the body's self-reported values remain valid, so they are given as usual. |

### Relative turning points

| baseline | concurrency | first tier where it is no longer faster (microseconds) | note |
|---|---|---|---|
| apscheduler | 1 | 40 | this is the smallest measured tier that satisfies the condition |
| apscheduler | 4 | 40 | this is the smallest measured tier that satisfies the condition |
| apscheduler | 16 | 40 | this is the smallest measured tier that satisfies the condition |
| apscheduler | 64 | 40 | this is the smallest measured tier that satisfies the condition |
| asyncio_pool | 1 | - | **not observed**: within the measured tiers this baseline is always faster than the framework under test |
| asyncio_pool | 4 | 10000 | this is the smallest measured tier that satisfies the condition |
| asyncio_pool | 16 | - | **not observed**: within the measured tiers this baseline is always faster than the framework under test |
| asyncio_pool | 64 | - | **not observed**: within the measured tiers this baseline is always faster than the framework under test |
| bare_thread | 1 | - | **not observed**: within the measured tiers this baseline is always faster than the framework under test |
| bare_thread | 4 | - | **not observed**: within the measured tiers this baseline is always faster than the framework under test |
| bare_thread | 16 | 40 | this is the smallest measured tier that satisfies the condition |
| bare_thread | 64 | - | **not observed**: within the measured tiers this baseline is always faster than the framework under test |
| process_pool | 1 | 40 | this is the smallest measured tier that satisfies the condition |
| process_pool | 4 | - | **not observed**: within the measured tiers this baseline is always faster than the framework under test |
| process_pool | 16 | - | **not observed**: within the measured tiers this baseline is always faster than the framework under test |
| process_pool | 64 | 40 | this is the smallest measured tier that satisfies the condition |
| thread_pool | 1 | - | **not observed**: within the measured tiers this baseline is always faster than the framework under test |
| thread_pool | 4 | - | **not observed**: within the measured tiers this baseline is always faster than the framework under test |
| thread_pool | 16 | - | **not observed**: within the measured tiers this baseline is always faster than the framework under test |
| thread_pool | 64 | - | **not observed**: within the measured tiers this baseline is always faster than the framework under test |

## Charts

![overhead_ratio](figures/en-overhead_ratio.svg)

> Body tiers at concurrency 4 only; ratios hold within a single run. The overhead numbers of concurrency 16, 64 were withheld (beyond this machine's parallelism; see the caliber section) and are not charted

![throughput](figures/en-throughput.svg)

> Body tier 40 microseconds only; throughput holds within a single run

![relative_multiple](figures/en-relative_multiple.svg)

> One panel per measured concurrency (4 in total); multiples hold within a single run

## Dimension: Latency percentiles and jitter

> end-to-end / concurrency; percentiles and relative spread are given in the same table  -  a metric that only gives the median is not acceptable

| concurrency | body tier (microseconds) | zoo | apscheduler | asyncio_pool | bare_thread | process_pool | thread_pool |
|---|---|---|---|---|---|---|---|
| 1 | 40 | 157.03 microseconds | 308.03 microseconds | 144.14 microseconds | 141.61 microseconds | 336.29 microseconds | 79.31 microseconds |
| 1 | 300 | 412.58 microseconds | 562.94 microseconds | 389.15 microseconds | 393.51 microseconds | 591.15 microseconds | 325.37 microseconds |
| 1 | 2700 | 2.778 milliseconds | 2.948 milliseconds | 2.663 milliseconds | 2.725 milliseconds | 3.013 milliseconds | 2.623 milliseconds |
| 1 | 10000 | 9.929 milliseconds | 10.112 milliseconds | 9.399 milliseconds | 9.630 milliseconds | 10.125 milliseconds | 9.369 milliseconds |
| 4 | 40 | 177.06 microseconds | 257.17 microseconds | 156.44 microseconds | 162.35 microseconds | 176.27 microseconds | 94.18 microseconds |
| 4 | 300 | 442.41 microseconds | 546.38 microseconds | 414.86 microseconds | 411.77 microseconds | 331.48 microseconds | 348.68 microseconds |
| 4 | 2700 | 2.878 milliseconds | 2.989 milliseconds | 2.729 milliseconds | 2.648 milliseconds | 1.400 milliseconds | 2.702 milliseconds |
| 4 | 10000 | 9.947 milliseconds | 10.095 milliseconds | 9.955 milliseconds | 9.618 milliseconds | 4.617 milliseconds | 9.839 milliseconds |
| 16 | 40 | 158.98 microseconds | 230.76 microseconds | 136.04 microseconds | 168.76 microseconds | 158.72 microseconds | 90.57 microseconds |
| 16 | 300 | 440.91 microseconds | 512.03 microseconds | 395.94 microseconds | 427.79 microseconds | 276.84 microseconds | 355.80 microseconds |
| 16 | 2700 | 2.885 milliseconds | 2.952 milliseconds | 2.725 milliseconds | 2.700 milliseconds | 1.370 milliseconds | 2.742 milliseconds |
| 16 | 10000 | 9.988 milliseconds | 10.097 milliseconds | 9.980 milliseconds | 9.717 milliseconds | 4.698 milliseconds | 9.867 milliseconds |
| 64 | 40 | 194.80 microseconds | 241.67 microseconds | 124.21 microseconds | 171.85 microseconds | 202.33 microseconds | 90.47 microseconds |
| 64 | 300 | 452.02 microseconds | 510.06 microseconds | 378.92 microseconds | 427.59 microseconds | 296.22 microseconds | 341.71 microseconds |
| 64 | 2700 | 2.928 milliseconds | 2.995 milliseconds | 2.779 milliseconds | 2.705 milliseconds | 1.357 milliseconds | 2.705 milliseconds |
| 64 | 10000 | 10.517 milliseconds | 10.128 milliseconds | 10.259 milliseconds | 9.718 milliseconds | 4.709 milliseconds | 10.415 milliseconds |

per-unit values (percentiles, spread, raw per-round samples) are in 'Appendix: all values' at the end.

## Dimension: Framework overhead share

> overhead = end-to-end / concurrency - the body duration measured within the same run (no cross-run subtraction). **cells with -**: that concurrency exceeds this machine's parallel capacity; the body's self-reported duration contains scheduling waits from oversubscription and is not comparable with the row's end-to-end per task, so no overhead figure is given (end-to-end and throughput live in their own dimensions)

| concurrency | body tier (microseconds) | zoo | apscheduler | asyncio_pool | bare_thread | process_pool | thread_pool |
|---|---|---|---|---|---|---|---|
| 1 | 40 | 68.65% | 81.31% | 67.95% | 63.22% | 83.79% | 39.28% |
| 1 | 300 | 27.69% | 46.45% | 26.86% | 24.74% | 47.86% | 11.59% |
| 1 | 2700 | 5.00% | 9.43% | 4.87% | 4.28% | 11.72% | 1.96% |
| 1 | 10000 | 1.74% | 4.37% | 1.50% | 1.20% | 3.66% | 0.60% |
| 4 | 40 | 72.38% | 77.32% | 66.81% | 68.24% | 46.48% | 48.94% |
| 4 | 300 | 31.63% | 42.90% | 25.61% | 29.06% | - | 15.26% |
| 4 | 2700 | 5.40% | 10.29% | 5.57% | 4.38% | - | 4.06% |
| 4 | 10000 | - | - | - | - | - | - |
| 16 | 40 | - | - | - | - | - | - |
| 16 | 300 | - | - | - | - | - | - |
| 16 | 2700 | - | - | - | - | - | - |
| 16 | 10000 | - | - | - | - | - | - |
| 64 | 40 | - | - | - | - | - | - |
| 64 | 300 | - | - | - | - | - | - |
| 64 | 2700 | - | - | - | - | - | - |
| 64 | 10000 | - | - | - | - | - | - |

per-unit values (percentiles, spread, raw per-round samples) are in 'Appendix: all values' at the end.

## Dimension: Throughput and concurrency scaling

> throughput = concurrency / median end-to-end; whether it keeps rising with concurrency is the scaling behaviour

| concurrency | body tier (microseconds) | zoo | apscheduler | asyncio_pool | bare_thread | process_pool | thread_pool |
|---|---|---|---|---|---|---|---|
| 1 | 40 | 6368.3 | 3246.5 | 6937.9 | 7061.4 | 2973.7 | 12609.0 |
| 1 | 300 | 2423.8 | 1776.4 | 2569.7 | 2541.2 | 1691.6 | 3073.4 |
| 1 | 2700 | 359.9 | 339.2 | 375.5 | 367.0 | 331.9 | 381.2 |
| 1 | 10000 | 100.7 | 98.9 | 106.4 | 103.8 | 98.8 | 106.7 |
| 4 | 40 | 22591.2 | 15554.0 | 25569.1 | 24638.4 | 22692.6 | 42474.1 |
| 4 | 300 | 9041.4 | 7321.0 | 9641.7 | 9714.1 | 12067.1 | 11472.0 |
| 4 | 2700 | 1389.7 | 1338.2 | 1465.6 | 1510.8 | 2856.9 | 1480.5 |
| 4 | 10000 | 402.1 | 396.3 | 401.8 | 415.9 | 866.4 | 406.6 |
| 16 | 40 | 100640.5 | 69336.7 | 117614.4 | 94809.1 | 100804.1 | 176656.6 |
| 16 | 300 | 36288.7 | 31248.4 | 40410.1 | 37401.2 | 57794.7 | 44969.3 |
| 16 | 2700 | 5545.8 | 5419.6 | 5872.3 | 5926.5 | 11681.1 | 5835.0 |
| 16 | 10000 | 1601.9 | 1584.6 | 1603.2 | 1646.6 | 3405.5 | 1621.6 |
| 64 | 40 | 328542.1 | 264822.7 | 515270.2 | 372428.2 | 316315.6 | 707442.1 |
| 64 | 300 | 141586.6 | 125476.2 | 168900.1 | 149676.6 | 216054.3 | 187293.3 |
| 64 | 2700 | 21859.0 | 21372.5 | 23028.2 | 23658.9 | 47176.6 | 23661.8 |
| 64 | 10000 | 6085.3 | 6319.2 | 6238.4 | 6585.6 | 13590.5 | 6144.8 |

per-unit values (percentiles, spread, raw per-round samples) are in 'Appendix: all values' at the end.

## Dimension: Overhead attribution

> scope: splits the per-task end-to-end into submit side / handoff / body / return and, at the same tier and concurrency, **compares each segment against every baseline**, answering 'which segment the excess falls on'. It is **diagnostic, not the main evidence**: it samples a small set of tiers, and does not compare total overhead against the baselines. **A negative handoff is not an error**: it means the submit call had not returned yet while the body was already running  -  schemes that count thread creation as submit side behave this way (measured: bare_thread handoff -343 microseconds). **The return segment includes waiting for the completion signal** (the last stretch before draining returns); it shares the main measurement's scope

**status: ok**

> this dimension is a **diagnosis**, not the main evidence: the sampled tiers and concurrencies are a fixed small batch, not the whole matrix. The concurrency ceiling is the machine's parallel capacity - beyond it, what is measured is contention;

sampled tiers (microseconds): 40, 10000; concurrency: [1, 4]; this machine's parallel capacity: 4

| scheme | concurrency | tier (microseconds) | submit side | handoff | body | return | status |
|---|---|---|---|---|---|---|---|
| apscheduler | 1 | 40 | 105.70 microseconds | 89.95 microseconds | 51.93 microseconds | 39.97 microseconds | ok |
| apscheduler | 4 | 40 | 88.92 microseconds | 519.14 microseconds | 58.30 microseconds | 235.85 microseconds | ok |
| apscheduler | 1 | 10000 | 190.14 microseconds | 131.76 microseconds | 9.755 milliseconds | 61.09 microseconds | ok |
| apscheduler | 4 | 10000 | 123.24 microseconds | 10.454 milliseconds | 14.656 milliseconds | 9.685 milliseconds | ok |
| asyncio_pool | 1 | 40 | 17.51 microseconds | 26.17 microseconds | 47.18 microseconds | 57.79 microseconds | ok |
| asyncio_pool | 4 | 40 | 19.75 microseconds | 261.66 microseconds | 53.31 microseconds | 295.26 microseconds | ok |
| asyncio_pool | 1 | 10000 | 25.81 microseconds | 39.08 microseconds | 9.477 milliseconds | 95.70 microseconds | ok |
| asyncio_pool | 4 | 10000 | 23.73 microseconds | 12.437 milliseconds | 14.392 milliseconds | 9.922 milliseconds | ok |
| bare_thread | 1 | 40 | 143.92 microseconds | -77.44 microseconds | 52.54 microseconds | 30.62 microseconds | ok |
| bare_thread | 4 | 40 | 153.77 microseconds | -75.41 microseconds | 51.61 microseconds | 337.75 microseconds | ok |
| bare_thread | 1 | 10000 | 5.201 milliseconds | -5092.36 microseconds | 9.693 milliseconds | 48.96 microseconds | ok |
| bare_thread | 4 | 10000 | 5.227 milliseconds | -5084.12 microseconds | 14.960 milliseconds | 9.964 milliseconds | ok |
| process_pool | 1 | 40 | - | - | - | - | not_measurable: the body's start/end timestamps are unavailable - it did not run in this process (cross-process units cannot be split this way) |
| process_pool | 4 | 40 | - | - | - | - | not_measurable: the body's start/end timestamps are unavailable - it did not run in this process (cross-process units cannot be split this way) |
| process_pool | 1 | 10000 | - | - | - | - | not_measurable: the body's start/end timestamps are unavailable - it did not run in this process (cross-process units cannot be split this way) |
| process_pool | 4 | 10000 | - | - | - | - | not_measurable: the body's start/end timestamps are unavailable - it did not run in this process (cross-process units cannot be split this way) |
| thread_pool | 1 | 40 | 8.44 microseconds | 10.47 microseconds | 44.63 microseconds | 15.66 microseconds | ok |
| thread_pool | 4 | 40 | 11.21 microseconds | 178.51 microseconds | 49.72 microseconds | 151.75 microseconds | ok |
| thread_pool | 1 | 10000 | 17.21 microseconds | 20.60 microseconds | 9.481 milliseconds | 40.80 microseconds | ok |
| thread_pool | 4 | 10000 | 13.85 microseconds | 10.088 milliseconds | 14.762 milliseconds | 9.041 milliseconds | ok |
| zoo | 1 | 40 | 64.50 microseconds | 41.26 microseconds | 56.29 microseconds | 73.36 microseconds | ok |
| zoo | 4 | 40 | 56.52 microseconds | 303.27 microseconds | 48.83 microseconds | 249.76 microseconds | ok |
| zoo | 1 | 10000 | 109.02 microseconds | 38.07 microseconds | 9.694 milliseconds | 140.97 microseconds | ok |
| zoo | 4 | 10000 | 86.25 microseconds | 10.271 milliseconds | 20.159 milliseconds | 5.744 milliseconds | ok |

### where the excess over each baseline falls

| baseline | concurrency | tier (microseconds) | excess per task | mostly on | gap in that segment |
|---|---|---|---|---|---|
| apscheduler | 1 | 40 | -57.52 microseconds | return | 33.39 microseconds |
| asyncio_pool | 1 | 40 | 87.51 microseconds | submit side | 46.99 microseconds |
| bare_thread | 1 | 40 | 85.04 microseconds | handoff | 118.70 microseconds |
| thread_pool | 1 | 40 | 156.15 microseconds | return | 57.70 microseconds |
| apscheduler | 4 | 40 | -246.88 microseconds | return | 13.91 microseconds |
| asyncio_pool | 4 | 40 | 16.03 microseconds | handoff | 41.61 microseconds |
| bare_thread | 4 | 40 | 200.23 microseconds | handoff | 378.67 microseconds |
| thread_pool | 4 | 40 | 279.47 microseconds | handoff | 124.76 microseconds |
| apscheduler | 1 | 10000 | -184.50 microseconds | return | 79.88 microseconds |
| asyncio_pool | 1 | 10000 | 324.83 microseconds | body | 217.01 microseconds |
| bare_thread | 1 | 10000 | 121.06 microseconds | handoff | 5.130 milliseconds |
| thread_pool | 1 | 10000 | 401.12 microseconds | body | 212.71 microseconds |
| apscheduler | 4 | 10000 | 1.529 milliseconds | body | 5.503 milliseconds |
| asyncio_pool | 4 | 10000 | 1.542 milliseconds | body | 5.767 milliseconds |
| bare_thread | 4 | 10000 | 8.639 milliseconds | handoff | 15.355 milliseconds |
| thread_pool | 4 | 10000 | 2.222 milliseconds | body | 5.396 milliseconds |

> instrumentation seams used (temporarily wrapped on the harness side, no framework code changed): BaseWaiter._dispatch_worker (handing a worker to the scheduling model); WorkerDispatchCore.resolve_period / resolve_phase / resolve_run_timeout (per-round policy lookups); BaseWaiter.execute_service (one scheduling round)

> **the cost of instrumentation itself** (on the framework-under-test side, paired from two rounds in the same subprocess, so it can be discounted honestly): zoo 40 microseconds / concurrency 1: without instrumentation 230.24 microseconds vs with 236.18 microseconds, a difference of 5.94 microseconds; zoo 40 microseconds / concurrency 4: without instrumentation 660.66 microseconds vs with 661.52 microseconds, a difference of 0.85 microseconds; zoo 10000 microseconds / concurrency 1: without instrumentation 9.995 milliseconds vs with 9.969 milliseconds, a difference of 26.32 microseconds (faster with instrumentation: this round's machine drift exceeded the instrumentation cost); zoo 10000 microseconds / concurrency 4: without instrumentation 41.278 milliseconds vs with 41.169 milliseconds, a difference of 109.45 microseconds (faster with instrumentation: this round's machine drift exceeded the instrumentation cost)

> **additivity of the sub-measurements**: the sub-measurements are **independent measurements**: each one's duration contains its own preemption waits, so their sum can exceed the submit side they all live in (about 3% measured). Do not read them as an additive partition.measured overlap this round: zoo 40 microseconds / concurrency 4 overlaps 3.5%; zoo 10000 microseconds / concurrency 1 overlaps 18.9%; zoo 10000 microseconds / concurrency 4 overlaps 94.3%

### what the submit side of the framework under test consists of

- concurrency 1, tier 40 microseconds: rest of the scheduling round (due checks, locks, list upkeep) 2.41 microseconds; dispatch (handed to the scheduling model, incl. pool enqueue) 28.54 microseconds; per-round policy lookups (period / phase / timeout each read config) 23.88 microseconds; submit side, other (the adapter's own bookkeeping and completion-signal state) 9.79 microseconds
- concurrency 4, tier 40 microseconds: rest of the scheduling round (due checks, locks, list upkeep) 0.00 microseconds; dispatch (handed to the scheduling model, incl. pool enqueue) 23.63 microseconds; per-round policy lookups (period / phase / timeout each read config) 24.83 microseconds; submit side, other (the adapter's own bookkeeping and completion-signal state) 4.82 microseconds
- concurrency 1, tier 10000 microseconds: rest of the scheduling round (due checks, locks, list upkeep) 1.43 microseconds; dispatch (handed to the scheduling model, incl. pool enqueue) 40.83 microseconds; per-round policy lookups (period / phase / timeout each read config) 47.10 microseconds; submit side, other (the adapter's own bookkeeping and completion-signal state) 18.34 microseconds
- concurrency 4, tier 10000 microseconds: rest of the scheduling round (due checks, locks, list upkeep) 0.00 microseconds; dispatch (handed to the scheduling model, incl. pool enqueue) 41.14 microseconds; per-round policy lookups (period / phase / timeout each read config) 44.81 microseconds; submit side, other (the adapter's own bookkeeping and completion-signal state) 8.34 microseconds

### Attribution conclusions

- At concurrency 1 and a 40 microsecond body, the framework under test is not slower overall than apscheduler (-58 microseconds per task): the two sides have different timing structures, and the per-segment differences have both signs, so no 'which segment' is given.
- At concurrency 1 and a 40 microsecond body, the framework under test spends 88 microseconds more per task than asyncio_pool, most of it on 'submit side' (+47 microseconds).
- At concurrency 1 and a 40 microsecond body, the framework under test spends 85 microseconds more per task than bare_thread, most of it on 'handoff' (+119 microseconds).
-   note: the per-segment differences in this group do not all share a sign  - 'submit side' is -79 microseconds, so read them together.
- At concurrency 1 and a 40 microsecond body, the framework under test spends 156 microseconds more per task than thread_pool, most of it on 'return' (+58 microseconds).
- In this group the framework's submit side consists of (microseconds): rest of the scheduling round (due checks, locks, list upkeep) 2, dispatch (handed to the scheduling model, incl. pool enqueue) 29, per-round policy lookups (period / phase / timeout each read config) 24, submit side, other (the adapter's own bookkeeping and completion-signal state) 10.
- At concurrency 4 and a 40 microsecond body, the framework under test is not slower overall than apscheduler (-247 microseconds per task): the two sides have different timing structures, and the per-segment differences have both signs, so no 'which segment' is given.
- At concurrency 4 and a 40 microsecond body, the framework under test spends 16 microseconds more per task than asyncio_pool, most of it on 'handoff' (+42 microseconds).
-   note: the per-segment differences in this group do not all share a sign  - 'return' is -46 microseconds, so read them together.
- At concurrency 4 and a 40 microsecond body, the framework under test spends 200 microseconds more per task than bare_thread, most of it on 'handoff' (+379 microseconds).
-   note: the per-segment differences in this group do not all share a sign  - 'submit side' is -97 microseconds, so read them together.
- At concurrency 4 and a 40 microsecond body, the framework under test spends 279 microseconds more per task than thread_pool, most of it on 'handoff' (+125 microseconds).
-   note: the per-segment differences in this group do not all share a sign  - 'body' is -1 microseconds, so read them together.
- In this group the framework's submit side consists of (microseconds): rest of the scheduling round (due checks, locks, list upkeep) 0, dispatch (handed to the scheduling model, incl. pool enqueue) 24, per-round policy lookups (period / phase / timeout each read config) 25, submit side, other (the adapter's own bookkeeping and completion-signal state) 5.
- At concurrency 1 and a 10000 microsecond body, the framework under test is not slower overall than apscheduler (-184 microseconds per task): the two sides have different timing structures, and the per-segment differences have both signs, so no 'which segment' is given.
- At concurrency 1 and a 10000 microsecond body, the framework under test spends 325 microseconds more per task than asyncio_pool, most of it on 'body' (+217 microseconds).
-   note: the per-segment differences in this group do not all share a sign  - 'handoff' is -1 microseconds, so read them together.
- At concurrency 1 and a 10000 microsecond body, the framework under test spends 121 microseconds more per task than bare_thread, most of it on 'handoff' (+5130 microseconds).
-   note: the per-segment differences in this group do not all share a sign  - 'submit side' is -5092 microseconds, so read them together.
- At concurrency 1 and a 10000 microsecond body, the framework under test spends 401 microseconds more per task than thread_pool, most of it on 'body' (+213 microseconds).
- In this group the framework's submit side consists of (microseconds): rest of the scheduling round (due checks, locks, list upkeep) 1, dispatch (handed to the scheduling model, incl. pool enqueue) 41, per-round policy lookups (period / phase / timeout each read config) 47, submit side, other (the adapter's own bookkeeping and completion-signal state) 18.
- At concurrency 4 and a 10000 microsecond body, the framework under test spends 1529 microseconds more per task than apscheduler, most of it on 'body' (+5503 microseconds).
-   note: the per-segment differences in this group do not all share a sign  - 'return' is -3941 microseconds, so read them together.
- At concurrency 4 and a 10000 microsecond body, the framework under test spends 1542 microseconds more per task than asyncio_pool, most of it on 'body' (+5767 microseconds).
-   note: the per-segment differences in this group do not all share a sign  - 'return' is -4177 microseconds, so read them together.
- At concurrency 4 and a 10000 microsecond body, the framework under test spends 8639 microseconds more per task than bare_thread, most of it on 'handoff' (+15355 microseconds).
-   note: the per-segment differences in this group do not all share a sign  - 'submit side' is -5141 microseconds, so read them together.
- At concurrency 4 and a 10000 microsecond body, the framework under test spends 2222 microseconds more per task than thread_pool, most of it on 'body' (+5396 microseconds).
-   note: the per-segment differences in this group do not all share a sign  - 'return' is -3297 microseconds, so read them together.
- In this group the framework's submit side consists of (microseconds): rest of the scheduling round (due checks, locks, list upkeep) 0, dispatch (handed to the scheduling model, incl. pool enqueue) 41, per-round policy lookups (period / phase / timeout each read config) 45, submit side, other (the adapter's own bookkeeping and completion-signal state) 8.

## Dimension: The cost of scheduling semantics

> scope: this dimension uses an **internal switch comparison** (same workload, semantics all off vs. enabled one by one) and **does not compare across schemes**  -  a bare writeup has no counterpart for priority / timeout / retry, so it can only answer 'what is this framework's semantics worth', not 'how much more expensive than a bare writeup'

**status: ok** (the conclusion was reached on the 'current' drive-surface generation installed during this measurement)

| semantics item | layer | how probed | measurable | evidence / reason |
|---|---|---|---|---|
| timeout | dispatch layer | behavioral probe (driven repeatedly through scheduling rounds; measures the moment the worker leaves the in-flight table) | yes | timeout reap observed: the worker left the in-flight table before the body ended (0.0122s) (body 0.15s, run_timeout 0.01s); judgment action at dispatch_core.py:472; what is in effect is observation and circuit-breaking (recording, marking unhealthy, removing the in-flight registration, stopping dispatch), **not termination** of the still-running worker - CPython cannot safely interrupt a running thread |
| priority | event layer | structural probe (checks whether the symbols required for execution exist) | no | not re-checked: this item's conclusion was reached on the 'previous' generation; taking it as this generation's fact asserts wiring in the report that the current generation does not have. Re-deriving it needs a probe design of its own (each item needs clean on/off states before it can be priced) |
| retry | event layer | structural probe (checks whether the symbols required for execution exist) | no | not re-checked: this item's conclusion was reached on the 'previous' generation; taking it as this generation's fact asserts wiring in the report that the current generation does not have. Re-deriving it needs a probe design of its own (each item needs clean on/off states before it can be priced) |

> this conclusion was reached on the version under test and its coordinates hold for that version only; after appending a version to matrix.yaml, re-run this probe

## Dimension: Native task line: three execution forms of the same subject

> scope: this line is measured on its own: it enters neither the main matrix's pivot tables nor its crossings, and its numbers must not be read against the baseline adapters listed there

**verdict: ok**

equivalence gate: passed, 15 inputs compared byte for byte against the Python reference (envelope 256 bytes)

> loaded task modbus_rtu.parse_response - Modbus RTU response-frame parsing: the input is a bare RTU ADU response frame (no transport framing), and the tier's frame bytes are its length

> tiers R1, R8, R125 (frame bytes 7, 21, 255); concurrencies 1, 4, 16, 64; each cell runs 100 warm-up plus 100 measured rounds, one child process per concurrency (4 in all)

crossover (read at the judged concurrency 1 only): the native form is faster on R8, R125; slower on R1; undecided on none. crossover point: up to 7 frame bytes the native form is not faster, at or above 21 bytes it is

> each figure is end-to-end per task (median of the measured rounds) - the same caliber as the main matrix's latency dimension

| tier (microseconds) | frame bytes | concurrency | pure Python | native extension | online routing | native extension / pure Python | online routing / pure Python | faster form |
|---|---|---|---|---|---|---|---|---|
| R1 | 7 | 1 | 191.86 microseconds | 197.11 microseconds | 239.64 microseconds | 1.03x | 1.25x | pure Python |
| R1 | 7 | 4 | 265.75 microseconds | 275.12 microseconds | 277.12 microseconds | 1.04x | 1.04x | pure Python |
| R1 | 7 | 16 | 258.25 microseconds | 270.49 microseconds | 278.69 microseconds | 1.05x | 1.08x | pure Python |
| R1 | 7 | 64 | 253.67 microseconds | 264.03 microseconds | 281.22 microseconds | 1.04x | 1.11x | pure Python |
| R8 | 21 | 1 | 216.95 microseconds | 205.05 microseconds | 261.69 microseconds | 0.95x | 1.21x | native extension |
| R8 | 21 | 4 | 295.66 microseconds | 289.60 microseconds | 301.84 microseconds | 0.98x | 1.02x | undecided |
| R8 | 21 | 16 | 289.49 microseconds | 283.11 microseconds | 306.98 microseconds | 0.98x | 1.06x | undecided |
| R8 | 21 | 64 | 277.33 microseconds | 275.01 microseconds | 306.46 microseconds | 0.99x | 1.11x | undecided |
| R125 | 255 | 1 | 642.48 microseconds | 433.90 microseconds | 487.97 microseconds | 0.68x | 0.76x | native extension |
| R125 | 255 | 4 | 708.90 microseconds | 501.54 microseconds | 514.68 microseconds | 0.71x | 0.73x | native extension |
| R125 | 255 | 16 | 692.00 microseconds | 501.51 microseconds | 513.39 microseconds | 0.72x | 0.74x | native extension |
| R125 | 255 | 64 | 670.87 microseconds | 494.37 microseconds | 518.43 microseconds | 0.74x | 0.77x | native extension |

> the routing verdict has two layers. Convergence: the arm actually taken is the majority of the arm counts inside the measured window, and the minority's share must be explainable by the exploration rate alone - counts are precise quantities, independent of wall time, and the per-round increments are archived so a reader can follow the settling. An arm the framework never took at all (cumulative snapshot, warm-up included) makes the majority the framework's default rather than a decision, and that is reported as such. Advantage: the two forms are contrasted by the paired per-task difference within one cycle (a sampled difference, not two separately measured medians subtracted), with no tolerance band (a band would turn 'inside it' into 'a tie'), and only when that difference's quartile range does not span zero - otherwise this caliber cannot tell them apart and the verdict is left undecided

routing at the judged concurrency, tier by tier (arm counts cover the measured window; the exploration rate is the framework's resolved value):

- R1 (7 bytes): arm counts native 1 / python 99; exploration rate 5.00% - converged to the python arm; agrees with the advantageous arm of the independent comparison (python arm)
- R8 (21 bytes): arm counts native 98 / python 2; exploration rate 5.00% - converged to the native arm; agrees with the advantageous arm of the independent comparison (native arm)
- R125 (255 bytes): arm counts native 98 / python 2; exploration rate 5.00% - converged to the native arm; agrees with the advantageous arm of the independent comparison (native arm)

> caliber statements for this dimension:

- this dimension contrasts three execution forms of the same framework (pure Python / native extension / online routing); the figures come from the scheduling dispatch layer, the same level as the main matrix's zoo adapter. Only end-to-end per task is obtainable here - on this path the framework reports no per-unit duration, and this repository adds no instrumentation to it
- the dispatch layer's own overhead is a cost common to all three forms, so it compresses the ratios between them: the framework's own per-arm records (see the arm caliber note) show a much wider gap than end-to-end does. Read these ratios as 'running this task through the dispatch layer', not as the execution form's own speedup
- the crossover verdict is read at the lowest concurrency only: beyond the machine's parallel capacity the per-task wall time contains queueing wait, and then whether native is faster no longer measures the execution form (the other concurrency tiers are still measured and archived)

## tiers where the framework under test is faster

> Tiers where the framework under test is ahead (the two sides differ by more than that tier's band). This comes from the **same within-run comparison** as the openly published unfavourable data  -  one side takes the faster tiers, the other the slower ones, so both cover the same ground and neither can pick its own tiers. **This section being empty is not a failure**: it should be empty when nothing is distinguishable or when the framework loses everywhere.

| baseline | concurrency | body tier (microseconds) | gap | median of the framework under test |
|---|---|---|---|---|
| apscheduler | 1 | 40 | the framework under test is 1.96x faster than apscheduler | 157.03 microseconds |
| process_pool | 1 | 40 | the framework under test is 2.14x faster than process_pool | 157.03 microseconds |
| apscheduler | 1 | 300 | the framework under test is 1.36x faster than apscheduler | 412.58 microseconds |
| process_pool | 1 | 300 | the framework under test is 1.43x faster than process_pool | 412.58 microseconds |
| apscheduler | 1 | 2700 | the framework under test is 1.06x faster than apscheduler | 2.778 milliseconds |
| process_pool | 1 | 2700 | the framework under test is 1.08x faster than process_pool | 2.778 milliseconds |
| apscheduler | 4 | 40 | the framework under test is 1.45x faster than apscheduler | 177.06 microseconds |
| apscheduler | 4 | 300 | the framework under test is 1.24x faster than apscheduler | 442.41 microseconds |
| apscheduler | 4 | 2700 | the framework under test is 1.04x faster than apscheduler | 2.878 milliseconds |
| apscheduler | 16 | 40 | the framework under test is 1.45x faster than apscheduler | 158.98 microseconds |
| apscheduler | 16 | 300 | the framework under test is 1.16x faster than apscheduler | 440.91 microseconds |
| apscheduler | 16 | 2700 | the framework under test is 1.02x faster than apscheduler | 2.885 milliseconds |
| apscheduler | 64 | 40 | the framework under test is 1.24x faster than apscheduler | 194.80 microseconds |
| apscheduler | 64 | 300 | the framework under test is 1.13x faster than apscheduler | 452.02 microseconds |
| apscheduler | 64 | 2700 | the framework under test is 1.02x faster than apscheduler | 2.928 milliseconds |

## Openly published unfavourable data

> Tiers where the framework under test is behind (the two sides differ by more than that tier's band). **A report missing this part is not acceptable**  -  a report that only shows where it wins gives the reader every reason to distrust it. Tiers whose difference is smaller than the band are not here but in the 'no winner' section: that is the verdict's outcome, not a place to hide things.

| baseline | concurrency | body tier (microseconds) | gap | median of the framework under test |
|---|---|---|---|---|
| thread_pool | 1 | 40 | thread_pool is 1.98x faster than zoo | 157.03 microseconds |
| thread_pool | 1 | 300 | thread_pool is 1.27x faster than zoo | 412.58 microseconds |
| asyncio_pool | 1 | 2700 | asyncio_pool is 1.04x faster than zoo | 2.778 milliseconds |
| thread_pool | 1 | 2700 | thread_pool is 1.06x faster than zoo | 2.778 milliseconds |
| asyncio_pool | 1 | 10000 | asyncio_pool is 1.06x faster than zoo | 9.929 milliseconds |
| bare_thread | 1 | 10000 | bare_thread is 1.03x faster than zoo | 9.929 milliseconds |
| thread_pool | 1 | 10000 | thread_pool is 1.06x faster than zoo | 9.929 milliseconds |
| thread_pool | 4 | 40 | thread_pool is 1.88x faster than zoo | 177.06 microseconds |
| bare_thread | 4 | 300 | bare_thread is 1.07x faster than zoo | 442.41 microseconds |
| process_pool | 4 | 300 | process_pool is 1.33x faster than zoo | 442.41 microseconds |
| thread_pool | 4 | 300 | thread_pool is 1.27x faster than zoo | 442.41 microseconds |
| asyncio_pool | 4 | 2700 | asyncio_pool is 1.05x faster than zoo | 2.878 milliseconds |
| bare_thread | 4 | 2700 | bare_thread is 1.09x faster than zoo | 2.878 milliseconds |
| process_pool | 4 | 2700 | process_pool is 2.06x faster than zoo | 2.878 milliseconds |
| thread_pool | 4 | 2700 | thread_pool is 1.07x faster than zoo | 2.878 milliseconds |
| bare_thread | 4 | 10000 | bare_thread is 1.03x faster than zoo | 9.947 milliseconds |
| process_pool | 4 | 10000 | process_pool is 2.15x faster than zoo | 9.947 milliseconds |
| asyncio_pool | 16 | 40 | asyncio_pool is 1.17x faster than zoo | 158.98 microseconds |
| thread_pool | 16 | 40 | thread_pool is 1.76x faster than zoo | 158.98 microseconds |
| asyncio_pool | 16 | 300 | asyncio_pool is 1.11x faster than zoo | 440.91 microseconds |
| process_pool | 16 | 300 | process_pool is 1.59x faster than zoo | 440.91 microseconds |
| thread_pool | 16 | 300 | thread_pool is 1.24x faster than zoo | 440.91 microseconds |
| asyncio_pool | 16 | 2700 | asyncio_pool is 1.06x faster than zoo | 2.885 milliseconds |
| bare_thread | 16 | 2700 | bare_thread is 1.07x faster than zoo | 2.885 milliseconds |
| process_pool | 16 | 2700 | process_pool is 2.11x faster than zoo | 2.885 milliseconds |
| thread_pool | 16 | 2700 | thread_pool is 1.05x faster than zoo | 2.885 milliseconds |
| bare_thread | 16 | 10000 | bare_thread is 1.03x faster than zoo | 9.988 milliseconds |
| process_pool | 16 | 10000 | process_pool is 2.13x faster than zoo | 9.988 milliseconds |
| asyncio_pool | 64 | 40 | asyncio_pool is 1.57x faster than zoo | 194.80 microseconds |
| bare_thread | 64 | 40 | bare_thread is 1.13x faster than zoo | 194.80 microseconds |
| thread_pool | 64 | 40 | thread_pool is 2.15x faster than zoo | 194.80 microseconds |
| asyncio_pool | 64 | 300 | asyncio_pool is 1.19x faster than zoo | 452.02 microseconds |
| bare_thread | 64 | 300 | bare_thread is 1.06x faster than zoo | 452.02 microseconds |
| process_pool | 64 | 300 | process_pool is 1.53x faster than zoo | 452.02 microseconds |
| thread_pool | 64 | 300 | thread_pool is 1.32x faster than zoo | 452.02 microseconds |
| asyncio_pool | 64 | 2700 | asyncio_pool is 1.05x faster than zoo | 2.928 milliseconds |
| bare_thread | 64 | 2700 | bare_thread is 1.08x faster than zoo | 2.928 milliseconds |
| process_pool | 64 | 2700 | process_pool is 2.16x faster than zoo | 2.928 milliseconds |
| thread_pool | 64 | 2700 | thread_pool is 1.08x faster than zoo | 2.928 milliseconds |
| bare_thread | 64 | 10000 | bare_thread is 1.08x faster than zoo | 10.517 milliseconds |
| process_pool | 64 | 10000 | process_pool is 2.23x faster than zoo | 10.517 milliseconds |

> the native-task line is a line of its own and is not mixed into the table above: at its judged concurrency 1 the native form is slower on R1 and the routing form is slower than pure interpretation on R1, R8; both are listed with equal prominence, from the same paired comparison of the same run, in the "Native task line: three execution forms of the same subject" section

## tiers with no winner

> Tiers where the two sides **cannot be told apart**: the difference is smaller than that tier's band. The band comes from the two sides' own across-round spread (stdev / median of the per-task medians) **within the same run**, so it varies with the sample and the reader can recompute it from the archived per-round samples. It also covers every measured tier and states a reason for each  -  **this section is not a place to hide things**: reporting noise-level differences as '1.00x faster' is.

> criterion: band = the framework's own across-round relative spread plus the baseline's; a single side's across-round relative spread = stdev / median of that unit's per-round per-task medians

band this run: 1.0% to 57.0% (it varies with each tier's spread).

| baseline | concurrency | body tier (microseconds) | ratio | reason |
|---|---|---|---|---|
| asyncio_pool | 1 | 40 | 0.92x | the two sides differ by 8.2%, less than the band of 23.0%  -  no winner |
| bare_thread | 1 | 40 | 0.90x | the two sides differ by 9.8%, less than the band of 23.8%  -  no winner |
| asyncio_pool | 1 | 300 | 0.94x | the two sides differ by 5.7%, less than the band of 7.2%  -  no winner |
| bare_thread | 1 | 300 | 0.95x | the two sides differ by 4.6%, less than the band of 10.1%  -  no winner |
| bare_thread | 1 | 2700 | 0.98x | the two sides differ by 1.9%, less than the band of 2.2%  -  no winner |
| apscheduler | 1 | 10000 | 1.02x | the two sides differ by 1.8%, less than the band of 1.9%  -  no winner |
| process_pool | 1 | 10000 | 1.02x | the two sides differ by 2.0%, less than the band of 3.7%  -  no winner |
| asyncio_pool | 4 | 40 | 0.88x | the two sides differ by 11.6%, less than the band of 20.2%  -  no winner |
| bare_thread | 4 | 40 | 0.92x | the two sides differ by 8.3%, less than the band of 17.7%  -  no winner |
| process_pool | 4 | 40 | 1.00x | the two sides differ by 0.4%, less than the band of 30.2%  -  no winner |
| asyncio_pool | 4 | 300 | 0.94x | the two sides differ by 6.2%, less than the band of 6.7%  -  no winner |
| apscheduler | 4 | 10000 | 1.01x | the two sides differ by 1.5%, less than the band of 2.0%  -  no winner |
| asyncio_pool | 4 | 10000 | 1.00x | the two sides differ by 0.1%, less than the band of 5.1%  -  no winner |
| thread_pool | 4 | 10000 | 0.99x | the two sides differ by 1.1%, less than the band of 1.9%  -  no winner |
| bare_thread | 16 | 40 | 1.06x | the two sides differ by 6.2%, less than the band of 7.3%  -  no winner |
| process_pool | 16 | 40 | 1.00x | the two sides differ by 0.2%, less than the band of 21.0%  -  no winner |
| bare_thread | 16 | 300 | 0.97x | the two sides differ by 3.0%, less than the band of 5.7%  -  no winner |
| apscheduler | 16 | 10000 | 1.01x | the two sides differ by 1.1%, less than the band of 1.2%  -  no winner |
| asyncio_pool | 16 | 10000 | 1.00x | the two sides differ by 0.1%, less than the band of 1.7%  -  no winner |
| thread_pool | 16 | 10000 | 0.99x | the two sides differ by 1.2%, less than the band of 2.9%  -  no winner |
| process_pool | 64 | 40 | 1.04x | the two sides differ by 3.9%, less than the band of 19.6%  -  no winner |
| apscheduler | 64 | 10000 | 0.96x | the two sides differ by 3.7%, less than the band of 5.1%  -  no winner |
| asyncio_pool | 64 | 10000 | 0.98x | the two sides differ by 2.5%, less than the band of 4.4%  -  no winner |
| thread_pool | 64 | 10000 | 0.99x | the two sides differ by 1.0%, less than the band of 3.9%  -  no winner |

## caliber limitations and sources of bias

- **measured layer**: dispatch layer: drives BaseWaiter.execute_service() directly (thread-pool mode) instead of Master's full run loop - the latter's service loop fires once per second, which would lift the latency scale from microseconds to seconds. **One scheduling round per submitted task**, so that whole round's bookkeeping is charged to this single task; Master's usage serves several workers per round, which would spread that cost over tasks
- **scope note** (apscheduler): APScheduler's scheduled-job model, including job registration and trigger-decision cost; the library swallows exceptions raised inside jobs, so this tier re-raises them during drain. **Caliber note**: the library offers no job-completion notification, so the completion signal comes from this tier's Condition, and the cost of one notification per task counts toward its end-to-end
- **scope note** (asyncio_pool): the asyncio event loop dispatches **synchronous** bodies (equivalent to loop.run_in_executor), not native coroutine concurrency; the difference from a thread pool is the event loop's own cost
- **scope note** (bare_thread): one threading.Thread per task, no pooling and no reuse; thread creation cost counts toward end-to-end. Bodies' self-reported timings are collected with a lock-free list.append (atomic under the GIL) - **instrumentation must not charge one tier a cost other tiers do not pay**, or the comparison becomes a comparison of instrumentation
- **scope note** (process_pool): ProcessPoolExecutor: pickling the body and its return value across processes counts toward end-to-end, so this tier is not directly comparable to the multi-threaded tiers
- **scope note** (thread_pool): concurrent.futures.ThreadPoolExecutor, pool size = concurrency; threads are reused, with no task-queue semantics
- **scope note** (zoo): end-to-end stops at the worker completing and being deregistered from the in-flight table; the framework's later result delivery splits by generation: the previous generation's delivery path is broken (dispatch looks reactors up by reactor_name rather than by topic, so every call raises), and the completion signal can only come from the framework's internal callback; the current generation has that fixed, and the completion signal comes from the framework's own result reactor. **Caliber note**: the completion signal comes from this tier's Condition, and the cost of one notification per task counts toward its end-to-end; a bias in that direction makes the framework under test look slower, i.e. it works against it rather than for it
- **groups with no overhead figure**: for the (adapter, concurrency) groups below, concurrency exceeds the parallel capacity of the machine this run was on: the body's self-reported duration contains scheduling waits from oversubscription, which is not the same thing as 'wall clock / concurrency', so subtracting them is meaningless (measured, it produced negative values and meaningless change factors). This report therefore **gives them no framework overhead and no crossing point**; end-to-end, throughput and the body's self-reported values remain valid and not one field has been dropped from the raw archive. The criterion is the logical core count from the environment self-report, plus one evidential backstop: a computed negative overhead is likewise treated as unreadable  -  a framework can only add time, never remove it
- **self-check items excluded from the verdict**: for these units the body-tier deviation did not take part in the verdict (concurrency above the lowest, or a tier so short that wall clock is dominated by scheduling granularity). The raw values are still in place, but the wall clock at those two places is not enough to conclude whether the calibration is right  -  **factor this in when reading those tiers' numbers**

> absolute durations: valid only within a single run; do not compare across runs

## bring your own baseline

the baselines of this harness are **open**: implement the four methods of `zoo_bench.adapters.BaseAdapter` (`setup` / `submit` / `drain` / `teardown`), register with `zoo_bench.adapters.registry.register`, or list your module path in `extra_modules` of `matrix.yaml`  -  **no file in this repository needs to change**. Your implementation goes through the same equivalence verification (submit N bodies, assert all ran and each exactly once); failing it means the data is not trusted. **We deliberately write this path into the report**: a report written by the framework's own maintainer, measuring that framework, gives every reason to suspect a slowed baseline or a chosen caliber; letting anyone submit their own baseline is what makes this report challengeable and falsifiable.

## self-check

- overall measurement self-check: passed
- process isolation: {'parent_pid': 71066, 'child_pids': [71522, 71525, 71528, 71531, 71534, 71540, 71546, 71552, 71558, 71576, 71594, 71612, 71630, 71687, 71745, 71800, 71866, 71868, 71870, 71872, 71874, 71879, 71884, 71890, 71895, 71912, 71929, 71946, 71963, 72025, 72087, 72152, 72221, 72255, 72289, 72323, 72357, 72490, 72623, 72756, 72889, 73418, 73947, 74476, 75006, 77119, 79232, 81345, 83460, 83464, 83468, 83472, 83476, 83483, 83490, 83497, 83504, 83523, 83542, 83561, 83580, 83647, 83714, 83781, 83852, 83854, 83856, 83858, 83860, 83865, 83870, 83875, 83880, 83897, 83914, 83931, 83948, 84004, 84061, 84120, 84187, 84189, 84191, 84193, 84195, 84200, 84205, 84210, 84215, 84232, 84251, 84268, 84285, 84350, 84415, 84480], 'all_distinct': True, 'none_equals_parent': True}

## Appendix: all values

> the body carries only pivot tables for readability; this appendix gives **every field as measured before any trimming** per unit (percentiles, spread, measured body durations, ...). **Not one value is dropped**: the details live inside the report body itself, not behind links, folds or attachments  -  'verifiable' that requires a separate lookup is not verifiable

### Appendix: Latency percentiles and jitter

| scheme | concurrency | body tier (microseconds) | median | p95 | p99 | relative spread |
|---|---|---|---|---|---|---|
| apscheduler | 1 | 40 | 308.03 microseconds | 349.96 microseconds | 358.97 microseconds | 0.066 |
| apscheduler | 1 | 300 | 562.94 microseconds | 604.91 microseconds | 627.65 microseconds | 0.046 |
| apscheduler | 1 | 2700 | 2.948 milliseconds | 3.005 milliseconds | 3.033 milliseconds | 0.010 |
| apscheduler | 1 | 10000 | 10.112 milliseconds | 10.352 milliseconds | 10.407 milliseconds | 0.012 |
| apscheduler | 4 | 40 | 257.17 microseconds | 287.55 microseconds | 303.15 microseconds | 0.056 |
| apscheduler | 4 | 300 | 546.38 microseconds | 561.11 microseconds | 576.20 microseconds | 0.022 |
| apscheduler | 4 | 2700 | 2.989 milliseconds | 3.032 milliseconds | 3.036 milliseconds | 0.009 |
| apscheduler | 4 | 10000 | 10.095 milliseconds | 10.212 milliseconds | 10.494 milliseconds | 0.011 |
| apscheduler | 16 | 40 | 230.76 microseconds | 268.29 microseconds | 271.09 microseconds | 0.069 |
| apscheduler | 16 | 300 | 512.03 microseconds | 538.09 microseconds | 543.38 microseconds | 0.024 |
| apscheduler | 16 | 2700 | 2.952 milliseconds | 2.978 milliseconds | 3.027 milliseconds | 0.009 |
| apscheduler | 16 | 10000 | 10.097 milliseconds | 10.192 milliseconds | 10.227 milliseconds | 0.006 |
| apscheduler | 64 | 40 | 241.67 microseconds | 262.14 microseconds | 262.49 microseconds | 0.054 |
| apscheduler | 64 | 300 | 510.06 microseconds | 530.38 microseconds | 538.66 microseconds | 0.027 |
| apscheduler | 64 | 2700 | 2.995 milliseconds | 3.033 milliseconds | 3.041 milliseconds | 0.007 |
| apscheduler | 64 | 10000 | 10.128 milliseconds | 10.588 milliseconds | 10.631 milliseconds | 0.024 |
| asyncio_pool | 1 | 40 | 144.14 microseconds | 173.18 microseconds | 187.49 microseconds | 0.089 |
| asyncio_pool | 1 | 300 | 389.15 microseconds | 418.60 microseconds | 424.26 microseconds | 0.036 |
| asyncio_pool | 1 | 2700 | 2.663 milliseconds | 2.700 milliseconds | 2.702 milliseconds | 0.005 |
| asyncio_pool | 1 | 10000 | 9.399 milliseconds | 9.941 milliseconds | 10.816 milliseconds | 0.036 |
| asyncio_pool | 4 | 40 | 156.44 microseconds | 177.87 microseconds | 184.77 microseconds | 0.080 |
| asyncio_pool | 4 | 300 | 414.86 microseconds | 432.37 microseconds | 438.97 microseconds | 0.023 |
| asyncio_pool | 4 | 2700 | 2.729 milliseconds | 2.771 milliseconds | 2.784 milliseconds | 0.009 |
| asyncio_pool | 4 | 10000 | 9.955 milliseconds | 10.372 milliseconds | 11.706 milliseconds | 0.042 |
| asyncio_pool | 16 | 40 | 136.04 microseconds | 147.86 microseconds | 158.86 microseconds | 0.061 |
| asyncio_pool | 16 | 300 | 395.94 microseconds | 412.78 microseconds | 419.97 microseconds | 0.024 |
| asyncio_pool | 16 | 2700 | 2.725 milliseconds | 2.760 milliseconds | 2.766 milliseconds | 0.009 |
| asyncio_pool | 16 | 10000 | 9.980 milliseconds | 10.087 milliseconds | 10.399 milliseconds | 0.012 |
| asyncio_pool | 64 | 40 | 124.21 microseconds | 127.64 microseconds | 128.05 microseconds | 0.018 |
| asyncio_pool | 64 | 300 | 378.92 microseconds | 383.97 microseconds | 386.20 microseconds | 0.008 |
| asyncio_pool | 64 | 2700 | 2.779 milliseconds | 2.808 milliseconds | 2.811 milliseconds | 0.008 |
| asyncio_pool | 64 | 10000 | 10.259 milliseconds | 10.474 milliseconds | 10.528 milliseconds | 0.018 |
| bare_thread | 1 | 40 | 141.61 microseconds | 176.49 microseconds | 185.39 microseconds | 0.097 |
| bare_thread | 1 | 300 | 393.51 microseconds | 447.16 microseconds | 485.24 microseconds | 0.066 |
| bare_thread | 1 | 2700 | 2.725 milliseconds | 2.778 milliseconds | 2.783 milliseconds | 0.012 |
| bare_thread | 1 | 10000 | 9.630 milliseconds | 9.751 milliseconds | 9.889 milliseconds | 0.008 |
| bare_thread | 4 | 40 | 162.35 microseconds | 181.13 microseconds | 187.59 microseconds | 0.055 |
| bare_thread | 4 | 300 | 411.77 microseconds | 429.43 microseconds | 433.87 microseconds | 0.023 |
| bare_thread | 4 | 2700 | 2.648 milliseconds | 2.675 milliseconds | 2.680 milliseconds | 0.005 |
| bare_thread | 4 | 10000 | 9.618 milliseconds | 9.680 milliseconds | 9.697 milliseconds | 0.004 |
| bare_thread | 16 | 40 | 168.76 microseconds | 176.90 microseconds | 178.35 microseconds | 0.023 |
| bare_thread | 16 | 300 | 427.79 microseconds | 433.37 microseconds | 434.78 microseconds | 0.007 |
| bare_thread | 16 | 2700 | 2.700 milliseconds | 2.718 milliseconds | 2.727 milliseconds | 0.004 |
| bare_thread | 16 | 10000 | 9.717 milliseconds | 9.795 milliseconds | 9.860 milliseconds | 0.004 |
| bare_thread | 64 | 40 | 171.85 microseconds | 175.73 microseconds | 176.05 microseconds | 0.014 |
| bare_thread | 64 | 300 | 427.59 microseconds | 431.98 microseconds | 433.04 microseconds | 0.006 |
| bare_thread | 64 | 2700 | 2.705 milliseconds | 2.724 milliseconds | 2.728 milliseconds | 0.004 |
| bare_thread | 64 | 10000 | 9.718 milliseconds | 9.769 milliseconds | 9.779 milliseconds | 0.004 |
| process_pool | 1 | 40 | 336.29 microseconds | 371.95 microseconds | 907.07 microseconds | 0.429 |
| process_pool | 1 | 300 | 591.15 microseconds | 634.03 microseconds | 1.158 milliseconds | 0.246 |
| process_pool | 1 | 2700 | 3.013 milliseconds | 3.061 milliseconds | 3.668 milliseconds | 0.055 |
| process_pool | 1 | 10000 | 10.125 milliseconds | 11.006 milliseconds | 11.208 milliseconds | 0.030 |
| process_pool | 4 | 40 | 176.27 microseconds | 244.14 microseconds | 252.09 microseconds | 0.181 |
| process_pool | 4 | 300 | 331.48 microseconds | 369.10 microseconds | 393.42 microseconds | 0.094 |
| process_pool | 4 | 2700 | 1.400 milliseconds | 1.518 milliseconds | 1.594 milliseconds | 0.041 |
| process_pool | 4 | 10000 | 4.617 milliseconds | 5.118 milliseconds | 5.600 milliseconds | 0.053 |
| process_pool | 16 | 40 | 158.72 microseconds | 204.66 microseconds | 211.37 microseconds | 0.160 |
| process_pool | 16 | 300 | 276.84 microseconds | 315.26 microseconds | 322.45 microseconds | 0.079 |
| process_pool | 16 | 2700 | 1.370 milliseconds | 1.469 milliseconds | 1.516 milliseconds | 0.032 |
| process_pool | 16 | 10000 | 4.698 milliseconds | 4.911 milliseconds | 4.934 milliseconds | 0.019 |
| process_pool | 64 | 40 | 202.33 microseconds | 237.75 microseconds | 243.81 microseconds | 0.102 |
| process_pool | 64 | 300 | 296.22 microseconds | 334.60 microseconds | 342.73 microseconds | 0.058 |
| process_pool | 64 | 2700 | 1.357 milliseconds | 1.410 milliseconds | 1.445 milliseconds | 0.019 |
| process_pool | 64 | 10000 | 4.709 milliseconds | 4.740 milliseconds | 4.866 milliseconds | 0.009 |
| thread_pool | 1 | 40 | 79.31 microseconds | 101.71 microseconds | 132.80 microseconds | 0.184 |
| thread_pool | 1 | 300 | 325.37 microseconds | 362.51 microseconds | 375.66 microseconds | 0.053 |
| thread_pool | 1 | 2700 | 2.623 milliseconds | 2.711 milliseconds | 2.937 milliseconds | 0.029 |
| thread_pool | 1 | 10000 | 9.369 milliseconds | 9.476 milliseconds | 9.510 milliseconds | 0.007 |
| thread_pool | 4 | 40 | 94.18 microseconds | 107.91 microseconds | 109.81 microseconds | 0.076 |
| thread_pool | 4 | 300 | 348.68 microseconds | 363.07 microseconds | 364.89 microseconds | 0.019 |
| thread_pool | 4 | 2700 | 2.702 milliseconds | 2.740 milliseconds | 2.761 milliseconds | 0.011 |
| thread_pool | 4 | 10000 | 9.839 milliseconds | 10.035 milliseconds | 10.105 milliseconds | 0.010 |
| thread_pool | 16 | 40 | 90.57 microseconds | 99.86 microseconds | 104.29 microseconds | 0.057 |
| thread_pool | 16 | 300 | 355.80 microseconds | 370.72 microseconds | 376.40 microseconds | 0.019 |
| thread_pool | 16 | 2700 | 2.742 milliseconds | 2.786 milliseconds | 2.800 milliseconds | 0.009 |
| thread_pool | 16 | 10000 | 9.867 milliseconds | 10.048 milliseconds | 10.787 milliseconds | 0.023 |
| thread_pool | 64 | 40 | 90.47 microseconds | 92.30 microseconds | 92.47 microseconds | 0.018 |
| thread_pool | 64 | 300 | 341.71 microseconds | 347.75 microseconds | 348.64 microseconds | 0.009 |
| thread_pool | 64 | 2700 | 2.705 milliseconds | 2.798 milliseconds | 2.946 milliseconds | 0.023 |
| thread_pool | 64 | 10000 | 10.415 milliseconds | 10.509 milliseconds | 10.544 milliseconds | 0.012 |
| zoo | 1 | 40 | 157.03 microseconds | 210.32 microseconds | 237.86 microseconds | 0.141 |
| zoo | 1 | 300 | 412.58 microseconds | 445.22 microseconds | 451.32 microseconds | 0.036 |
| zoo | 1 | 2700 | 2.778 milliseconds | 2.829 milliseconds | 2.873 milliseconds | 0.010 |
| zoo | 1 | 10000 | 9.929 milliseconds | 10.006 milliseconds | 10.039 milliseconds | 0.007 |
| zoo | 4 | 40 | 177.06 microseconds | 230.68 microseconds | 233.14 microseconds | 0.122 |
| zoo | 4 | 300 | 442.41 microseconds | 486.15 microseconds | 486.91 microseconds | 0.044 |
| zoo | 4 | 2700 | 2.878 milliseconds | 2.957 milliseconds | 2.975 milliseconds | 0.017 |
| zoo | 4 | 10000 | 9.947 milliseconds | 10.070 milliseconds | 10.087 milliseconds | 0.009 |
| zoo | 16 | 40 | 158.98 microseconds | 172.71 microseconds | 180.55 microseconds | 0.050 |
| zoo | 16 | 300 | 440.91 microseconds | 485.64 microseconds | 488.09 microseconds | 0.050 |
| zoo | 16 | 2700 | 2.885 milliseconds | 2.933 milliseconds | 2.957 milliseconds | 0.011 |
| zoo | 16 | 10000 | 9.988 milliseconds | 10.064 milliseconds | 10.066 milliseconds | 0.006 |
| zoo | 64 | 40 | 194.80 microseconds | 218.06 microseconds | 219.26 microseconds | 0.095 |
| zoo | 64 | 300 | 452.02 microseconds | 463.04 microseconds | 471.99 microseconds | 0.022 |
| zoo | 64 | 2700 | 2.928 milliseconds | 3.004 milliseconds | 3.061 milliseconds | 0.015 |
| zoo | 64 | 10000 | 10.517 milliseconds | 10.925 milliseconds | 11.084 milliseconds | 0.027 |

### Appendix: Framework overhead share

| scheme | concurrency | body tier (microseconds) | framework overhead | overhead share | body as measured |
|---|---|---|---|---|---|
| apscheduler | 1 | 40 | 250.47 microseconds | 81.31% | 57.56 microseconds |
| apscheduler | 1 | 300 | 261.46 microseconds | 46.45% | 301.48 microseconds |
| apscheduler | 1 | 2700 | 277.96 microseconds | 9.43% | 2.670 milliseconds |
| apscheduler | 1 | 10000 | 442.08 microseconds | 4.37% | 9.670 milliseconds |
| apscheduler | 4 | 40 | 198.84 microseconds | 77.32% | 58.33 microseconds |
| apscheduler | 4 | 300 | 234.40 microseconds | 42.90% | 311.98 microseconds |
| apscheduler | 4 | 2700 | 307.46 microseconds | 10.29% | 2.682 milliseconds |
| apscheduler | 4 | 10000 | - | - | 14.587 milliseconds |
| apscheduler | 16 | 40 | - | - | 53.43 microseconds |
| apscheduler | 16 | 300 | - | - | 312.27 microseconds |
| apscheduler | 16 | 2700 | - | - | 2.685 milliseconds |
| apscheduler | 16 | 10000 | - | - | 18.654 milliseconds |
| apscheduler | 64 | 40 | - | - | 51.87 microseconds |
| apscheduler | 64 | 300 | - | - | 295.36 microseconds |
| apscheduler | 64 | 2700 | - | - | 2.740 milliseconds |
| apscheduler | 64 | 10000 | - | - | 19.471 milliseconds |
| asyncio_pool | 1 | 40 | 97.94 microseconds | 67.95% | 46.20 microseconds |
| asyncio_pool | 1 | 300 | 104.53 microseconds | 26.86% | 284.63 microseconds |
| asyncio_pool | 1 | 2700 | 129.72 microseconds | 4.87% | 2.534 milliseconds |
| asyncio_pool | 1 | 10000 | 140.82 microseconds | 1.50% | 9.258 milliseconds |
| asyncio_pool | 4 | 40 | 104.52 microseconds | 66.81% | 51.92 microseconds |
| asyncio_pool | 4 | 300 | 106.23 microseconds | 25.61% | 308.64 microseconds |
| asyncio_pool | 4 | 2700 | 152.12 microseconds | 5.57% | 2.577 milliseconds |
| asyncio_pool | 4 | 10000 | - | - | 14.730 milliseconds |
| asyncio_pool | 16 | 40 | - | - | 53.92 microseconds |
| asyncio_pool | 16 | 300 | - | - | 304.81 microseconds |
| asyncio_pool | 16 | 2700 | - | - | 2.605 milliseconds |
| asyncio_pool | 16 | 10000 | - | - | 20.437 milliseconds |
| asyncio_pool | 64 | 40 | - | - | 52.81 microseconds |
| asyncio_pool | 64 | 300 | - | - | 299.54 microseconds |
| asyncio_pool | 64 | 2700 | - | - | 2.649 milliseconds |
| asyncio_pool | 64 | 10000 | - | - | 59.077 milliseconds |
| bare_thread | 1 | 40 | 89.53 microseconds | 63.22% | 52.09 microseconds |
| bare_thread | 1 | 300 | 97.34 microseconds | 24.74% | 296.17 microseconds |
| bare_thread | 1 | 2700 | 116.72 microseconds | 4.28% | 2.608 milliseconds |
| bare_thread | 1 | 10000 | 115.64 microseconds | 1.20% | 9.514 milliseconds |
| bare_thread | 4 | 40 | 110.79 microseconds | 68.24% | 51.56 microseconds |
| bare_thread | 4 | 300 | 119.68 microseconds | 29.06% | 292.09 microseconds |
| bare_thread | 4 | 2700 | 115.99 microseconds | 4.38% | 2.532 milliseconds |
| bare_thread | 4 | 10000 | - | - | 14.615 milliseconds |
| bare_thread | 16 | 40 | - | - | 50.83 microseconds |
| bare_thread | 16 | 300 | - | - | 294.67 microseconds |
| bare_thread | 16 | 2700 | - | - | 2.563 milliseconds |
| bare_thread | 16 | 10000 | - | - | 14.718 milliseconds |
| bare_thread | 64 | 40 | - | - | 50.76 microseconds |
| bare_thread | 64 | 300 | - | - | 295.36 microseconds |
| bare_thread | 64 | 2700 | - | - | 2.565 milliseconds |
| bare_thread | 64 | 10000 | - | - | 14.787 milliseconds |
| process_pool | 1 | 40 | 281.77 microseconds | 83.79% | 54.52 microseconds |
| process_pool | 1 | 300 | 282.90 microseconds | 47.86% | 308.25 microseconds |
| process_pool | 1 | 2700 | 353.01 microseconds | 11.72% | 2.660 milliseconds |
| process_pool | 1 | 10000 | 370.20 microseconds | 3.66% | 9.755 milliseconds |
| process_pool | 4 | 40 | 81.93 microseconds | 46.48% | 94.34 microseconds |
| process_pool | 4 | 300 | - | - | 571.62 microseconds |
| process_pool | 4 | 2700 | - | - | 4.823 milliseconds |
| process_pool | 4 | 10000 | - | - | 17.617 milliseconds |
| process_pool | 16 | 40 | - | - | 90.76 microseconds |
| process_pool | 16 | 300 | - | - | 588.78 microseconds |
| process_pool | 16 | 2700 | - | - | 9.835 milliseconds |
| process_pool | 16 | 10000 | - | - | 52.389 milliseconds |
| process_pool | 64 | 40 | - | - | 89.98 microseconds |
| process_pool | 64 | 300 | - | - | 575.97 microseconds |
| process_pool | 64 | 2700 | - | - | 27.524 milliseconds |
| process_pool | 64 | 10000 | - | - | 167.164 milliseconds |
| thread_pool | 1 | 40 | 31.15 microseconds | 39.28% | 48.16 microseconds |
| thread_pool | 1 | 300 | 37.71 microseconds | 11.59% | 287.66 microseconds |
| thread_pool | 1 | 2700 | 51.42 microseconds | 1.96% | 2.572 milliseconds |
| thread_pool | 1 | 10000 | 56.20 microseconds | 0.60% | 9.313 milliseconds |
| thread_pool | 4 | 40 | 46.09 microseconds | 48.94% | 48.09 microseconds |
| thread_pool | 4 | 300 | 53.22 microseconds | 15.26% | 295.45 microseconds |
| thread_pool | 4 | 2700 | 109.72 microseconds | 4.06% | 2.592 milliseconds |
| thread_pool | 4 | 10000 | - | - | 19.244 milliseconds |
| thread_pool | 16 | 40 | - | - | 48.03 microseconds |
| thread_pool | 16 | 300 | - | - | 305.97 microseconds |
| thread_pool | 16 | 2700 | - | - | 2.660 milliseconds |
| thread_pool | 16 | 10000 | - | - | 20.155 milliseconds |
| thread_pool | 64 | 40 | - | - | 47.00 microseconds |
| thread_pool | 64 | 300 | - | - | 285.50 microseconds |
| thread_pool | 64 | 2700 | - | - | 2.624 milliseconds |
| thread_pool | 64 | 10000 | - | - | 76.944 milliseconds |
| zoo | 1 | 40 | 107.81 microseconds | 68.65% | 49.22 microseconds |
| zoo | 1 | 300 | 114.25 microseconds | 27.69% | 298.32 microseconds |
| zoo | 1 | 2700 | 139.00 microseconds | 5.00% | 2.639 milliseconds |
| zoo | 1 | 10000 | 173.17 microseconds | 1.74% | 9.756 milliseconds |
| zoo | 4 | 40 | 128.15 microseconds | 72.38% | 48.91 microseconds |
| zoo | 4 | 300 | 139.93 microseconds | 31.63% | 302.48 microseconds |
| zoo | 4 | 2700 | 155.55 microseconds | 5.40% | 2.723 milliseconds |
| zoo | 4 | 10000 | - | - | 14.996 milliseconds |
| zoo | 16 | 40 | - | - | 42.93 microseconds |
| zoo | 16 | 300 | - | - | 302.36 microseconds |
| zoo | 16 | 2700 | - | - | 2.737 milliseconds |
| zoo | 16 | 10000 | - | - | 25.176 milliseconds |
| zoo | 64 | 40 | - | - | 47.76 microseconds |
| zoo | 64 | 300 | - | - | 284.11 microseconds |
| zoo | 64 | 2700 | - | - | 2.747 milliseconds |
| zoo | 64 | 10000 | - | - | 114.364 milliseconds |

### Appendix: Throughput and concurrency scaling

| scheme | concurrency | body tier (microseconds) | throughput (tasks/s) |
|---|---|---|---|
| apscheduler | 1 | 40 | 3246.5 |
| apscheduler | 1 | 300 | 1776.4 |
| apscheduler | 1 | 2700 | 339.2 |
| apscheduler | 1 | 10000 | 98.9 |
| apscheduler | 4 | 40 | 15554.0 |
| apscheduler | 4 | 300 | 7321.0 |
| apscheduler | 4 | 2700 | 1338.2 |
| apscheduler | 4 | 10000 | 396.3 |
| apscheduler | 16 | 40 | 69336.7 |
| apscheduler | 16 | 300 | 31248.4 |
| apscheduler | 16 | 2700 | 5419.6 |
| apscheduler | 16 | 10000 | 1584.6 |
| apscheduler | 64 | 40 | 264822.7 |
| apscheduler | 64 | 300 | 125476.2 |
| apscheduler | 64 | 2700 | 21372.5 |
| apscheduler | 64 | 10000 | 6319.2 |
| asyncio_pool | 1 | 40 | 6937.9 |
| asyncio_pool | 1 | 300 | 2569.7 |
| asyncio_pool | 1 | 2700 | 375.5 |
| asyncio_pool | 1 | 10000 | 106.4 |
| asyncio_pool | 4 | 40 | 25569.1 |
| asyncio_pool | 4 | 300 | 9641.7 |
| asyncio_pool | 4 | 2700 | 1465.6 |
| asyncio_pool | 4 | 10000 | 401.8 |
| asyncio_pool | 16 | 40 | 117614.4 |
| asyncio_pool | 16 | 300 | 40410.1 |
| asyncio_pool | 16 | 2700 | 5872.3 |
| asyncio_pool | 16 | 10000 | 1603.2 |
| asyncio_pool | 64 | 40 | 515270.2 |
| asyncio_pool | 64 | 300 | 168900.1 |
| asyncio_pool | 64 | 2700 | 23028.2 |
| asyncio_pool | 64 | 10000 | 6238.4 |
| bare_thread | 1 | 40 | 7061.4 |
| bare_thread | 1 | 300 | 2541.2 |
| bare_thread | 1 | 2700 | 367.0 |
| bare_thread | 1 | 10000 | 103.8 |
| bare_thread | 4 | 40 | 24638.4 |
| bare_thread | 4 | 300 | 9714.1 |
| bare_thread | 4 | 2700 | 1510.8 |
| bare_thread | 4 | 10000 | 415.9 |
| bare_thread | 16 | 40 | 94809.1 |
| bare_thread | 16 | 300 | 37401.2 |
| bare_thread | 16 | 2700 | 5926.5 |
| bare_thread | 16 | 10000 | 1646.6 |
| bare_thread | 64 | 40 | 372428.2 |
| bare_thread | 64 | 300 | 149676.6 |
| bare_thread | 64 | 2700 | 23658.9 |
| bare_thread | 64 | 10000 | 6585.6 |
| process_pool | 1 | 40 | 2973.7 |
| process_pool | 1 | 300 | 1691.6 |
| process_pool | 1 | 2700 | 331.9 |
| process_pool | 1 | 10000 | 98.8 |
| process_pool | 4 | 40 | 22692.6 |
| process_pool | 4 | 300 | 12067.1 |
| process_pool | 4 | 2700 | 2856.9 |
| process_pool | 4 | 10000 | 866.4 |
| process_pool | 16 | 40 | 100804.1 |
| process_pool | 16 | 300 | 57794.7 |
| process_pool | 16 | 2700 | 11681.1 |
| process_pool | 16 | 10000 | 3405.5 |
| process_pool | 64 | 40 | 316315.6 |
| process_pool | 64 | 300 | 216054.3 |
| process_pool | 64 | 2700 | 47176.6 |
| process_pool | 64 | 10000 | 13590.5 |
| thread_pool | 1 | 40 | 12609.0 |
| thread_pool | 1 | 300 | 3073.4 |
| thread_pool | 1 | 2700 | 381.2 |
| thread_pool | 1 | 10000 | 106.7 |
| thread_pool | 4 | 40 | 42474.1 |
| thread_pool | 4 | 300 | 11472.0 |
| thread_pool | 4 | 2700 | 1480.5 |
| thread_pool | 4 | 10000 | 406.6 |
| thread_pool | 16 | 40 | 176656.6 |
| thread_pool | 16 | 300 | 44969.3 |
| thread_pool | 16 | 2700 | 5835.0 |
| thread_pool | 16 | 10000 | 1621.6 |
| thread_pool | 64 | 40 | 707442.1 |
| thread_pool | 64 | 300 | 187293.3 |
| thread_pool | 64 | 2700 | 23661.8 |
| thread_pool | 64 | 10000 | 6144.8 |
| zoo | 1 | 40 | 6368.3 |
| zoo | 1 | 300 | 2423.8 |
| zoo | 1 | 2700 | 359.9 |
| zoo | 1 | 10000 | 100.7 |
| zoo | 4 | 40 | 22591.2 |
| zoo | 4 | 300 | 9041.4 |
| zoo | 4 | 2700 | 1389.7 |
| zoo | 4 | 10000 | 402.1 |
| zoo | 16 | 40 | 100640.5 |
| zoo | 16 | 300 | 36288.7 |
| zoo | 16 | 2700 | 5545.8 |
| zoo | 16 | 10000 | 1601.9 |
| zoo | 64 | 40 | 328542.1 |
| zoo | 64 | 300 | 141586.6 |
| zoo | 64 | 2700 | 21859.0 |
| zoo | 64 | 10000 | 6085.3 |