Add more benchmarks #11

saeta · 2020-05-13T20:38:42Z

No description provided.

Sources/Benchmarks/NonBlockingThreadPool.swift

Before: name time std iterations --------------------------------------------------------------------------------------------------------------------- NonBlockingThreadPool: join, one level 690.0 ns ± 163522.04859994436 76147 NonBlockingThreadPool: join, two levels 2064.0 ns ± 316859.10411173315 15613 NonBlockingThreadPool: join, three levels 4941.0 ns ± 351192.3836080291 11462 NonBlockingThreadPool: join, four levels, three on thread pool thread 6369.0 ns ± 400867.0104087665 8426 NonBlockingThreadPool: parallel for, one level 2862068.0 ns ± 395500.70262353314 509 NonBlockingThreadPool: parallel for, two levels 2609616.5 ns ± 251679.60080760892 486 NonBlockingThreadPool: parallel for, one level, grain size 10 345706.0 ns ± 247728.9607701072 2859 NonBlockingThreadPool: parallel for, one level, grain size 100 142316.0 ns ± 292455.2283264145 4417 NonBlockingThreadPool: parallel for, one level, grain size 1000 78977.0 ns ± 214052.89967520352 6321 NonBlockingThreadPool: parallel for, one level, grain size 2000 50729.0 ns ± 164390.21843606833 7247 NonBlockingThreadPool: parallel for, one level, grain size 5000 39656.0 ns ± 190176.9821963313 8649 NonBlockingThreadPool: parallel for, two levels, grain size 10 & 100 1657159.5 ns ± 806196.7324534839 842 NonBlockingThreadPool: dispatch concurrent perform, one level 237865.0 ns ± 21098.405428758302 6223 NonBlockingThreadPool: dispatch concurrent perform, two levels 40180.0 ns ± 5290.912573146778 38178 NonBlockingThreadPool: sequential one level 370.0 ns ± 224.22220630175732 1000000 NonBlockingThreadPool: sequential two levels 485.0 ns ± 279.54687747682283 1000000 After: name time std iterations --------------------------------------------------------------------------------------------------------------------- NonBlockingThreadPool: join, one level 484.0 ns ± 161129.49945627034 78655 NonBlockingThreadPool: join, two levels 1446.0 ns ± 255823.17176138048 29165 NonBlockingThreadPool: join, three levels 3369.0 ns ± 244942.08954270577 21109 NonBlockingThreadPool: join, four levels, three on thread pool thread 3833.0 ns ± 270662.97199023387 15008 NonBlockingThreadPool: parallel for, one level 2661959.0 ns ± 360213.96138696355 543 NonBlockingThreadPool: parallel for, two levels 2469215.5 ns ± 334972.7168585615 538 NonBlockingThreadPool: parallel for, one level, grain size 10 327907.0 ns ± 473609.1412896736 2708 NonBlockingThreadPool: parallel for, one level, grain size 100 125174.5 ns ± 329418.62320991984 3934 NonBlockingThreadPool: parallel for, one level, grain size 1000 73999.0 ns ± 238626.6115465731 6940 NonBlockingThreadPool: parallel for, one level, grain size 2000 50771.0 ns ± 205223.8350497277 7089 NonBlockingThreadPool: parallel for, one level, grain size 5000 26571.5 ns ± 220587.47147580027 8398 NonBlockingThreadPool: parallel for, two levels, grain size 10 & 100 1162189.0 ns ± 940733.3616910677 991 NonBlockingThreadPool: dispatch concurrent perform, one level 247549.0 ns ± 23605.70739546625 5815 NonBlockingThreadPool: dispatch concurrent perform, two levels 41744.0 ns ± 6449.1291628621575 33719 NonBlockingThreadPool: sequential one level 402.0 ns ± 170.32164151542221 1000000 NonBlockingThreadPool: sequential two levels 466.0 ns ± 195.13074674728864 1000000

dabrahams · 2020-05-22T03:35:27Z

Sources/PenguinParallel/NonblockingThreadPool/NonBlockingThreadPool.swift

      } else {
        wakeupWorkerIfRequired()
      }
    } else {
      // Called not from within the threadpool; pick a victim thread from the pool at random.
      // TODO: use a faster RNG!
      let victim = Int.random(in: 0..<queues.count)
-      if let bounced = queues[victim].pushBack(fn) {


Thought you were against victim. It's not API, but still? Personally, I have no objections, but why waste an opportunity to give you a hard time?

dabrahams · 2020-05-22T03:36:18Z

Sources/PenguinParallel/NonblockingThreadPool/NonBlockingThreadPool.swift

+        } else {
+          wakeupWorkerIfRequired()
+
+          // Execute local half using recurion.


Suggested change

// Execute local half using recurion.

// Execute local half using recursion.

dabrahams

It would be nice if you'd add some comments describing what you're trying to measure and why. That would make it possible for me to evaluate the code without guessing at intent, and would provide guideposts for future benchmarks

dabrahams · 2020-05-22T03:43:24Z

Sources/PenguinParallel/NonblockingThreadPool/NonBlockingThreadPool.swift

+  }
+
+  fileprivate func executeJoinTask(_ task: UnsafeMutablePointer<JoinTask>) {
+    assert(!task.pointee.isDoneAcquiring(), "Join task done before even starting execution?!?")


Do you want to use assert in benchmarking code, given that it will be compiled away in release mode?

Hmm, not quite sure I follow? I do intentionally want to use an assert because I do want it compiled away in release mode. But having these asserts sprinkled throughout the code has been helpful in ensuring the implementation is correct during testing.

That's fine! I was just asking.

dabrahams · 2020-05-22T03:45:08Z

It looks like a lot of what happens in this PR is not just adding benchmarking. Can it be broken into smaller units or at least more completely described?

saeta

I'm going to move this into draft form. I'm not ready for it to go in. Thanks for the comments! I appreciate them.

saeta · 2020-05-22T15:24:17Z

Sources/PenguinParallel/NonblockingThreadPool/NonBlockingThreadPool.swift

      } else {
        wakeupWorkerIfRequired()
      }
    } else {
      // Called not from within the threadpool; pick a victim thread from the pool at random.
      // TODO: use a faster RNG!
      let victim = Int.random(in: 0..<queues.count)
-      if let bounced = queues[victim].pushBack(fn) {


saeta · 2020-05-22T15:25:14Z

Sources/PenguinParallel/NonblockingThreadPool/NonBlockingThreadPool.swift

+  }
+
+  fileprivate func executeJoinTask(_ task: UnsafeMutablePointer<JoinTask>) {
+    assert(!task.pointee.isDoneAcquiring(), "Join task done before even starting execution?!?")


Hmm, not quite sure I follow? I do intentionally want to use an assert because I do want it compiled away in release mode. But having these asserts sprinkled throughout the code has been helpful in ensuring the implementation is correct during testing.

saeta added 3 commits May 13, 2020 13:37

Add benchmarks and reduce time spent on heap operations in parallelFor.

b03e1ac

Remove .iterations from Benchmarks, as the auto-tuning works better.

5ed0864

Remove last explicit iterations setting.

ba46a39

shabalind reviewed May 13, 2020

View reviewed changes

Sources/Benchmarks/NonBlockingThreadPool.swift Outdated Show resolved Hide resolved

saeta added 2 commits May 13, 2020 13:42

Update Benchmark package.

9ddfef4

dabrahams reviewed May 22, 2020

View reviewed changes

saeta commented May 22, 2020

View reviewed changes

saeta marked this pull request as draft May 22, 2020 15:26

This was referenced May 22, 2020

ThreadPool cleanup (3/n): Switch to vectorized API & remove unused/co… #32

Merged

Improve the ComputeThreadPool abstraction #33

Open

shabalind self-requested a review November 18, 2020 18:27

texasmichelle changed the base branch from master to main December 8, 2020 01:30

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add more benchmarks #11

Add more benchmarks #11

saeta commented May 13, 2020

dabrahams May 22, 2020

saeta May 22, 2020

dabrahams May 22, 2020

dabrahams left a comment

dabrahams May 22, 2020

saeta May 22, 2020

dabrahams May 22, 2020

dabrahams commented May 22, 2020

saeta left a comment

saeta May 22, 2020

saeta May 22, 2020

	// Execute local half using recurion.
	// Execute local half using recursion.

Add more benchmarks #11

Are you sure you want to change the base?

Add more benchmarks #11

Conversation

saeta commented May 13, 2020

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

dabrahams left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

dabrahams commented May 22, 2020

saeta left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment