[Vm-dev] Code Compilation | Performance regression in recent VMs

[Eliot Miranda]

Hi Marcel,

On Fri, Aug 20, 2021 at 3:22 AM Marcel Taeumel <marcel.taeumel at hpi.de>
wrote:

>
> Hi all --
>
> I found the culprit. The regression started with the introduction of
> *FastCPrimitiveFlag* in VMMaker.oscog-eem.2995. That new feature needs
> more adjustments, I suppose. :-)
>
> My previous numbers were kind of wrong. I compared cog.spur against
> stack.spur. The latter being, in comparison, blazingly fast in #compileAll.
> Something is wrong with current cog.spur, at least on Windows (64-bit).
>
> CommitHash: d8576b0 Plugins: 2021*0820*0011
>
> stack.pur
> Morph methodDict size. " 1493 "
> [Morph compileAll] timeToRun " 1293 "
>
> cog.spur
> Morph methodDict size. " 1493 "
> [Morph compileAll] timeToRun " 23018 "
>
> Here are more tests with cog.spur:
>
> CommitHash: 128d14994 Plugins: 2021*0505*0056
> cog.spur -- [Morph compileAll] timeToRun " 1294 "
>
> CommitHash: 9d4e82529 Plugins: 2021*0712*1800
> cog.spur -- [Morph compileAll] timeToRun " 910 "
>
> CommitHash: c0d1a5574 Plugins: 2021*0721*1710 -- VMMaker.oscog-eem.2991
> cog.spur -- [Morph compileAll] timeToRun " 901 "
>
> CommitHash: 118a02c69 Plugins: 2021*0723*2247
> cog.spur -- [Morph compileAll] timeToRun " 984 "
>
> *(BUILD BROKEN  -- VMMaker.oscog-eem.2995)*
> *(VM CRASHES between VMMaker.oscog-eem.2996 and VMMaker.oscog-eem.3009)*
>
> CommitHash: 03c4ef31b Plugins: 2021*0801*0511 -- VMMaker.oscog-eem.3009
> cog.spur -- [Morph compileAll] timeToRun " 20787 "
>
> ***
>
> Note that #compileAll does not write into the .changes file.
>
> Note that the last "good" commit hash
> is 118a02c69f3e4c8c29e8ba867a78bd39f9a53cb4 (still using
> VMMaker.oscog-eem.2991). I added the tag "last-before-FastCPrimitiveFlag"
> to this commit.
>

2992 is good too.  I see the regresion.  It's huge; an 11 x slow down.
Looking at VM parameters I see a 4 fold change in the number of check event
calls, but that can be explained by the longer run time, so I don't think
this is significant.  ANyway, looking closely now.

>
> Now let's review and fix this. ^__^
>
> Best,
> Marcel
>
> Am 19.08.2021 15:55:34 schrieb Marcel Taeumel <marcel.taeumel at hpi.de>:
> [Morph compileAll] timeToRunWithoutGC
>
> VM 202108160631 : 24607 milliseconds
> VM 202108010511 : 1078 milliseconds
>
> Best,
> Marcel
>
> Am 19.08.2021 15:40:03 schrieb Marcel Taeumel <marcel.taeumel at hpi.de>:
> Hi all!
>
> There is a serious performance regression in recent VM (202108160631) regarding
> code compilation. I could only narrow it down to CompiledMethod >>
> #methodClass:. Still investigating. The start-up is also very slow. Maybe
> GC-related?
>
> This regression was not there in 202108010511.
>
> Best,
> Marcel
>
> _,,,^..^,,,_
best, Eliot
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.squeakfoundation.org/pipermail/vm-dev/attachments/20210820/fc58ea93/attachment.html>