rt(threaded): basic self-tuning of injection queue #5720

carllerche · 2023-05-24T23:26:12Z

Each multi-threaded runtime worker prioritizes pulling tasks off of its local queue. Every so often, it checks the injection (global) queue for work submitted there. Previously, "every so often," was a constant "number of tasks polled" value. Tokio sets a default of 61, but allows users to configure this value.

If workers are under load with tasks that are slow to poll, the injection queue can be starved. To prevent starvation in this case, this commit implements some basic self-tuning. The multi-threaded scheduler tracks the mean task poll time using an exponentially-weighted moving average. It then uses this value to pick an interval at which to check the injection queue.

This commit is a first pass at adding self-tuning to the scheduler. There are other values in the scheduler that could benefit from self-tuning (e.g. the maintenance interval). Additionally, the current-thread scheduler could also benfit from self-tuning. However, we have reached the point where we should start investigating ways to unify logic in both schedulers. Adding self-tuning to the current-thread scheduler will be punted until after this unification.

With this change, I can now add the benchmark mentioned in #5712 and this brings it down from 2s -> 100ms about.

Each multi-threaded runtime worker prioritizes pulling tasks off of its local queue. Every so often, it checks the injection (global) queue for work submitted there. Previously, "every so often," was a constant "number of tasks polled" value. Tokio sets a default of 61, but allows users to configure this value. If workers are under load with tasks that are slow to poll, the injection queue can be starved. To prevent starvation in this case, this commit implements some basic self-tuning. The multi-threaded scheduler tracks the mean task poll time using an exponentially-weighted moving average. It then uses this value to pick an interval at which to check the injection queue. This commit is a first pass at adding self-tuning to the scheduler. There are other values in the scheduler that could benefit from self-tuning (e.g. the maintenance interval). Additionally, the current-thread scheduler could also benfit from self-tuning. However, we have reached the point where we should start investigating ways to unify logic in both schedulers. Adding self-tuning to the current-thread scheduler will be punted until after this unification.

carllerche · 2023-05-24T23:28:20Z

tokio/src/runtime/scheduler/multi_thread/worker.rs

-                // was called, turning the poll into a "blocking op". In this
-                // case, we don't want to measure the poll time as it doesn't
-                // really count as an async poll anymore.
-                core.metrics.end_poll();


These changes might be a bit controversial, but when we poll the LIFO slot, we aren't "really" polling a new task, but we are batching polling the lifo task under the initially polled task.

i think this feels reasonable. can we add a comment explaining that rationale?

carllerche · 2023-05-24T23:41:59Z

I will work on fixing CI.

I will also investigate tests, but I think that might be hard because they are timing based.

Noah-Kennedy · 2023-05-25T19:55:11Z

We should probably test this change on a few different workloads to see how it performs.

carllerche · 2023-05-25T20:14:02Z

@Noah-Kennedy Go for it. Tokio's benchmarks (except the one this targets) remain unchanged (margin of error).

This reverts commit 9ff9218.

…tuning2

hawkw

the implementation looks really solid! i had a few minor nitpicks but nothing blocking.

benches/rt_multi_threaded.rs

tokio/src/runtime/builder.rs

hawkw · 2023-05-29T18:16:20Z

tokio/src/runtime/scheduler/multi_thread/worker.rs

-                // was called, turning the poll into a "blocking op". In this
-                // case, we don't want to measure the poll time as it doesn't
-                // really count as an async poll anymore.
-                core.metrics.end_poll();


i think this feels reasonable. can we add a comment explaining that rationale?

tokio/src/runtime/scheduler/multi_thread/stats.rs

tokio/src/runtime/scheduler/multi_thread/worker.rs

tokio/tests/rt_threaded.rs

PR #5720 introduced runtime self-tuning. It included a test that attempts to verify self-tuning logic. The test is heavily reliant on timing details. This patch attempts to make the test a bit more reliable by not assuming tuning will converge within a set amount of time.

This PR contains the following updates: | Package | Type | Update | Change | |---|---|---|---| | [tokio](https://tokio.rs) ([source](https://github.com/tokio-rs/tokio)) | dependencies | minor | `1.28.2` -> `1.29.1` | | [tokio](https://tokio.rs) ([source](https://github.com/tokio-rs/tokio)) | dev-dependencies | minor | `1.28.2` -> `1.29.1` | --- ### Release Notes <details> <summary>tokio-rs/tokio (tokio)</summary> ### [`v1.29.1`](https://github.com/tokio-rs/tokio/releases/tag/tokio-1.29.1): Tokio v1.29.1 [Compare Source](tokio-rs/tokio@tokio-1.29.0...tokio-1.29.1) ##### Fixed - rt: fix nesting two `block_in_place` with a `block_on` between (#5837]) #5837]: tokio-rs/tokio#5837 ### [`v1.29.0`](https://github.com/tokio-rs/tokio/releases/tag/tokio-1.29.0): Tokio v1.29.0 [Compare Source](tokio-rs/tokio@tokio-1.28.2...tokio-1.29.0) Technically a breaking change, the `Send` implementation is removed from `runtime::EnterGuard`. This change fixes a bug and should not impact most users. ##### Breaking - rt: `EnterGuard` should not be `Send` (#5766]) ##### Fixed - fs: reduce blocking ops in `fs::read_dir` (#5653]) - rt: fix possible starvation (#5686], #5712]) - rt: fix stacked borrows issue in `JoinSet` (#5693]) - rt: panic if `EnterGuard` dropped incorrect order (#5772]) - time: do not overflow to signal value (#5710]) - fs: wait for in-flight ops before cloning `File` (#5803]) ##### Changed - rt: reduce time to poll tasks scheduled from outside the runtime (#5705], #5720]) ##### Added - net: add uds doc alias for unix sockets (#5659]) - rt: add metric for number of tasks (#5628]) - sync: implement more traits for channel errors (#5666]) - net: add nodelay methods on TcpSocket (#5672]) - sync: add `broadcast::Receiver::blocking_recv` (#5690]) - process: add `raw_arg` method to `Command` (#5704]) - io: support PRIORITY epoll events (#5566]) - task: add `JoinSet::poll_join_next` (#5721]) - net: add support for Redox OS (#5790]) ##### Unstable - rt: add the ability to dump task backtraces (#5608], #5676], #5708], #5717]) - rt: instrument task poll times with a histogram (#5685]) #5766]: tokio-rs/tokio#5766 #5653]: tokio-rs/tokio#5653 #5686]: tokio-rs/tokio#5686 #5712]: tokio-rs/tokio#5712 #5693]: tokio-rs/tokio#5693 #5772]: tokio-rs/tokio#5772 #5710]: tokio-rs/tokio#5710 #5803]: tokio-rs/tokio#5803 #5705]: tokio-rs/tokio#5705 #5720]: tokio-rs/tokio#5720 #5659]: tokio-rs/tokio#5659 #5628]: tokio-rs/tokio#5628 #5666]: tokio-rs/tokio#5666 #5672]: tokio-rs/tokio#5672 #5690]: tokio-rs/tokio#5690 #5704]: tokio-rs/tokio#5704 #5566]: tokio-rs/tokio#5566 #5721]: tokio-rs/tokio#5721 #5790]: tokio-rs/tokio#5790 #5608]: tokio-rs/tokio#5608 #5676]: tokio-rs/tokio#5676 #5708]: tokio-rs/tokio#5708 #5717]: tokio-rs/tokio#5717 #5685]: tokio-rs/tokio#5685 </details> --- ### Configuration 📅 **Schedule**: Branch creation - At any time (no schedule defined), Automerge - At any time (no schedule defined). 🚦 **Automerge**: Disabled by config. Please merge this manually once you are satisfied. ♻ **Rebasing**: Whenever PR becomes conflicted, or you tick the rebase/retry checkbox. 🔕 **Ignore**: Close this PR and you won't be reminded about these updates again. --- - [ ] If you want to rebase/retry this PR, check this box --- This PR has been generated by [Renovate Bot](https://github.com/renovatebot/renovate).  Co-authored-by: cabr2-bot <cabr2.help@gmail.com> Reviewed-on: https://codeberg.org/Calciumdibromid/CaBr2/pulls/1958 Reviewed-by: crapStone <crapstone01@gmail.com> Co-authored-by: Calciumdibromid Bot <cabr2_bot@noreply.codeberg.org> Co-committed-by: Calciumdibromid Bot <cabr2_bot@noreply.codeberg.org>

carllerche added A-tokio Area: The main tokio crate M-runtime Module: tokio/runtime T-performance Topic: performance and benchmarks labels May 24, 2023

github-actions bot added the R-loom Run loom tests on this PR label May 24, 2023

carllerche force-pushed the rt-inject-interval-tuning2 branch from 9b49b59 to df96c16 Compare May 24, 2023 23:27

carllerche commented May 24, 2023

View reviewed changes

carllerche mentioned this pull request May 24, 2023

rt(threaded): basic self-tuning of injection queue #5715

Closed

carllerche added 6 commits May 25, 2023 08:18

fix loom tests

b7af661

fix build on MSRV

e93894b

fix clippy

987d081

avoid tuning in loom tests

6db32b2

add a test

2e59745

try increasing loom scope

9ff9218

hawkw self-requested a review May 25, 2023 19:28

carllerche added 12 commits May 25, 2023 13:15

Revert "try increasing loom scope"

073412b

This reverts commit 9ff9218.

disable tuning test when running cross tests

d07794e

fix warnings

266a11d

fmt

d91022d

tweak test

034096b

try again

0dc2b8f

try again

e3ab179

fix constant to match what it was before

4350c47

Merge remote-tracking branch 'origin/master' into rt-inject-interval-…

616a9ca

…tuning2

try again

0bb5765

try again

eb4a7d3

tweak ci

b9eee6a

Merge remote-tracking branch 'origin/master' into rt-inject-interval-…

c1784a9

…tuning2

hawkw approved these changes May 29, 2023

View reviewed changes

carllerche added 4 commits May 30, 2023 12:50

apply feedback

1bb42a9

tweak tuning

c5d5fe9

try again

5d4c19f

try again

da2a153

Noah-Kennedy approved these changes May 31, 2023

View reviewed changes

Merge branch 'master' into rt-inject-interval-tuning2

297099e

Darksonn approved these changes Jun 1, 2023

View reviewed changes

carllerche merged commit 79a7e78 into master Jun 1, 2023

carllerche deleted the rt-inject-interval-tuning2 branch June 1, 2023 15:13

carllerche mentioned this pull request Jun 1, 2023

ci: fix spurious CI failure #5752

Merged

aliu mentioned this pull request Nov 3, 2023

worker_total_busy_duration seems to be calculated incorrectly when using multi-threaded runtime #6129

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

rt(threaded): basic self-tuning of injection queue #5720

rt(threaded): basic self-tuning of injection queue #5720

carllerche commented May 24, 2023

carllerche May 24, 2023

hawkw May 29, 2023

carllerche May 30, 2023

carllerche commented May 24, 2023

Noah-Kennedy commented May 25, 2023

carllerche commented May 25, 2023

hawkw left a comment

hawkw May 29, 2023

rt(threaded): basic self-tuning of injection queue #5720

rt(threaded): basic self-tuning of injection queue #5720

Conversation

carllerche commented May 24, 2023

carllerche May 24, 2023

Choose a reason for hiding this comment

hawkw May 29, 2023

Choose a reason for hiding this comment

carllerche May 30, 2023

Choose a reason for hiding this comment

carllerche commented May 24, 2023

Noah-Kennedy commented May 25, 2023

carllerche commented May 25, 2023

hawkw left a comment

Choose a reason for hiding this comment

hawkw May 29, 2023

Choose a reason for hiding this comment