Locking Hopac threads

[Horusiath]

I'm writing here, since I don't have a place to ask a question.

I'm trying to use Hopac within a library in hope to get some performance boost over existing work done with Async (link to PR). However moving to Hopac from Async turned to even more performance degradation. After some profiling I've managed to get some snapshot from sample executing program:

It looks like Hopac-managed threads (top 4 on the image) are waiting idle for 70-80% of their time - ManualResetEventSlim.Wait and Monitor.Wait are probably invoked from Hopac itself. What can cause that?

The code I'm using, consists only of Job operations, no channels, ivars or alternatives, no explicit locks:

• Job.result
• Job.raises
• Job.conCollect
• Job.conIgnore
• Job.seqIgnore (however it's not used in the execution path, the picture reffers to)
• Async.Global.ofJob to expose job as async to public API

Any help will be appreciated.

[polytypic]

FYI, I'm currently quite busy at work and otherwise so it might take some time before I have time to look into this properly.

[polytypic]

Build of FSharp.Data.GraphQL seems to be failing at the moment. How can the project be built and the benchmark run?

[Horusiath]

It's strange - build should be possible from dev branch. The image I've attached has been made using DotTrace Performance Profiler on sample console application. Should I share full example somewhere to reproduce the problem?

[polytypic]

I did a git clone and checked out dev and tried to run build.sh from OS X, build.cmd from Windows, and tried to build Release from VS 2015. Both build scripts produced a ton of warnings and errors and ultimately reported failure. Build from VS also failed. So, what is the correct way to build it?

[Horusiath]

@polytypic could you try again with the latest version. It should work now.

[polytypic]

FYI, I managed to compile and run the benchmarks. So, I'll take a closer look at the code when I get time. Based on my previous brief look at the code, it seemed like you were basically using Hopac for parallelism rather than async and that there might not be very large amounts of parallel jobs to execute. If so, then in such a case one needs to be careful to get the jobs running on many worker threads.

BTW, doc generation seems to fails when you try to build from a fresh clone. Also, the benchmarks do work on Windows+.Net, but do not work on OS X+Mono.