ndmitchell
High CPU use with WSL
ghcid on Windows was constantly using 25% CPU (which I guess is all of one core but I'm not certain) even when there were no changes to any files that it's watching. Is there anything in particular that I should be tweaking?
$ ~/.cabal/bin/ghcid --version
Auto reloading GHCi daemon v0.6.6
$ ghc --version
The Glorious Glasgow Haskell Compilation System, version 7.10.3
Weird. I can only imagine it's the file watching code going nuts, or GC making a freak out, but with only 1.2Mb active mem GC seems unlikely. Can you reproduce it?
It seems to happen every time I launch ghcid on this particular project. I am using 7.10 though so it may be something that's not longer a problem for people who are living in the present day.
Can you launch with --verbose? Can you launch with +RTS -S? If you can profile that's even better. Trying with a ghcid compiled with a newer GHC, even if the project remains on 7.10, would be interesting.
With +RTS -S I get lots of lines like the below. Does that help?
522240 496 358568 0.000 0.000 3.844 4.817 0 0 (Gen: 0)
522240 496 358568 0.000 0.000 3.859 4.825 0 0 (Gen: 0)
522240 496 358568 0.000 0.000 3.859 4.832 0 0 (Gen: 0)
522240 496 358568 0.000 0.000 3.875 4.838 0 0 (Gen: 0)
522240 496 358568 0.000 0.000 3.875 4.844 0 0 (Gen: 0)
522240 496 358568 0.000 0.000 3.891 4.856 0 0 (Gen: 0)
522240 496 358568 0.000 0.000 3.891 4.867 0 0 (Gen: 0)
522240 496 358568 0.000 0.000 3.906 4.874 0 0 (Gen: 0)
522240 496 358568 0.000 0.000 3.906 4.881 0 0 (Gen: 0)
522240 496 358568 0.000 0.000 3.922 4.887 0 0 (Gen: 0)
522240 496 358568 0.000 0.000 3.922 4.897 0 0 (Gen: 0)
522240 496 358568 0.000 0.000 3.938 4.906 0 0 (Gen: 0)
522240 496 358568 0.000 0.000 3.938 4.912 0 0 (Gen: 0)
I think those incrementing numbers are seconds so it looks to be in the order of 100 per second. Unfortunately it seems I haven't compiled in support for -I0. If I get a moment to recompile I'll try it.
I realised this is under WSL rather than a Windows native binary. The plot thickens.
I have also ran into this issue under WSL. Trying to run with +RTS -I0 gives ghcid: Most RTS options are disabled. Link with -rtsopts to enable them. So I would need to build ghcid from scratch for this to try it seems. I wonder if there may be a nix one-liner snippet to do so quickly.
Btw, may be relevant: I have observed this issue when using WSL GHCid with NTFS (/mnt/c/...), haven't tried the 'more native' lxss/ext file system approach (/home/...).
I've recompiled GHCid with -rtsopts from master and ran it with +RTS -I0; no discernible change in CPU usage.
I've also tried +RTS -S, and like @tomjaguarpaw, I get many of those (maybe dozens per second) once GHCi finished loading (only sparsely interspersed between loading messages before that). +RTS -I0 -S Doesn't seem to change that.
1044480 368 1794320 0.000 0.000 57.375 320.851 0 0 (Gen: 0)
1044480 368 1794320 0.000 0.000 57.375 320.859 0 0 (Gen: 0)
1044480 368 1794320 0.000 0.000 57.375 320.870 0 0 (Gen: 0)
1044480 368 1794320 0.000 0.000 57.375 320.892 0 0 (Gen: 0)
1044480 368 1794320 0.000 0.000 57.391 320.908 0 0 (Gen: 0)
1044480 368 1794320 0.000 0.000 57.391 320.915 0 0 (Gen: 0)
1044480 368 1794320 0.000 0.000 57.391 320.942 0 0 (Gen: 0)
1044480 368 1794320 0.000 0.000 57.391 320.951 0 0 (Gen: 0)
1044480 368 1794320 0.000 0.000 57.406 320.976 0 0 (Gen: 0)
1044480 368 1794320 0.000 0.000 57.406 320.981 0 0 (Gen: 0)
1044480 368 1794320 0.000 0.000 57.406 320.994 0 0 (Gen: 0)
1044480 368 1794320 0.000 0.000 57.406 321.002 0 0 (Gen: 0)
1044480 368 1794320 0.000 0.000 57.406 321.012 0 0 (Gen: 0)
1044480 368 1794320 0.000 0.000 57.406 321.026 0 0 (Gen: 0)
1044480 368 1794320 0.000 0.000 57.422 321.047 0 0 (Gen: 0)
Here is the last summary I got after ctrl+c with -I0 -S:
25,681,470,392 bytes allocated in the heap
16,876,008 bytes copied during GC
1,639,056 bytes maximum residency (4 sample(s))
73,072 bytes maximum slop
5 MB total memory in use (0 MB lost due to fragmentation)
Tot time (elapsed) Avg pause Max pause
Gen 0 24584 colls, 0 par 0.672s 1.106s 0.0000s 0.0075s
Gen 1 4 colls, 0 par 0.016s 0.015s 0.0037s 0.0087s
TASKS: 6 (1 bound, 5 peak workers (5 total), using -N1)
SPARKS: 0 (0 converted, 0 overflowed, 0 dud, 0 GC'd, 0 fizzled)
INIT time 0.000s ( 0.002s elapsed)
MUT time 56.797s (321.214s elapsed)
GC time 0.688s ( 1.121s elapsed)
EXIT time 0.000s ( 0.008s elapsed)
Total time 57.484s (322.345s elapsed)
Alloc rate 452,163,440 bytes per MUT second
Productivity 98.8% of total user, 99.7% of total elapsed
gc_alloc_block_sync: 0
whitehole_spin: 0
gen[0].sync: 0
gen[1].sync: 0
Is that helpful? Any next idea @ndmitchell or @tomjaguarpaw?
Is that helpful? Any next idea @ndmitchell or @tomjaguarpaw?
I'm afraid I haven't made any progress debugging this and I don't know where to start!
The fact most of the time is in MUT suggests if you compiled with profiling and ran it might tell you what where the time was going in the profile. I'd recommend that as the next step.
@tomjaguarpaw - are you on WSL? Windows?
@Wizek - does this happen for all projects or just one? Which GHC version?
On Wed, Aug 29, 2018 at 03:27:09PM -0700, Neil Mitchell wrote:
@tomjaguarpaw - are you on WSL? Windows?
Yes, I'm on WSL.
See https://github.com/ndmitchell/ghcid/issues/121#issuecomment-417964326 - seems a more recent WSL has improved things. It also reminds me that if you are having issues, --poll might be your friend.
I can quickly comment that I dimly remember having observed the same high CPU use with --poll as well. Maybe I can try --poll 100 to verify that the CPU use is independent from file watching.
I'm also glad to read that this may be improved in an windows insider build, I'll attempt to verify that.
Just checked: --poll=100 still uses 25% CPU constantly. And as expected it is taking a long time to pick up on file changes, so polling seems to be in effect indeed.
Profiling done: https://gist.github.com/Wizek/f520108a5282aefb9333bca081fd565d
Does this give you any ideas? Next up: might be trying the insider build.
Verified, insider build 17744 works well! Tried 2 combinations: editing files from a WSL editor vs a win32 editor; both on NTFS, both without poll, both picked up file changes and both work without excess CPU use. Wonderful!
For the sake of my curiosity, if one of you can tell me if you can make any sense of my attached prof file, I would appreciate knowing what was going wrong in the stable builds. Also, it may take months before this fix lands on stable.
Unfortunately that profile is garbage, it says that it spent 100% of the time executing main which is <built in>. I've never seen anything like that before. Suggests something has upset the runtime quite substantially.
doesn't that just mean that the profiled build is lacking the -fprof-auto (or related flags) in the right places (the right deps)?
Probably @lspitzner - I confess I've never tried generating a profile without setting the flags first. My standard flags are -prof -auto-all -caf-all.