Yes, and we have to do several passes to construct the model with references (so we store a mapping table). Maybe a collection or two are GCed. Vincent -----Original Message----- From: Pharo-dev [mailto:pharo-dev-bounces@lists.pharo.org] On Behalf Of stepharo Sent: dimanche 24 janvier 2016 13:40 To: Pharo Development List Subject: Re: [Pharo-dev] Spur Garbage collection takes 4 times more when loading MSE (WAS: mse loading looks slower :() Thanks for this analysis. This is really interesting. I wonder what is the type of the garbage collected object. In MSE we have mainly strings. Stef Le 24/1/16 12:43, Vincent BLONDEAU a écrit :

Hi,

I made the benchmarks with the files you provided. I have more or less the same magnitude: Version 504: 0:00:01:09.021 Version 1175: 0:00:02:37.507

However, by launching it in the time profiler (MooseModel new importFromMSEStream: (StandardFileStream readOnlyFileNamed: 'd:/ArgoUML-0-34.mse')), it takes 504: 1 min 55 1175: 4 min 25 Well there is a delta...

After investigation, the standard process has almost the same duration (120 secs for prespur and 140 secs for spur). But, there is a large difference in GC time:

504: not spur **Memory** old +144,822,000 bytes young -8,293,660 bytes used +136,528,340 bytes free -104,186,788 bytes

**GCs** full 1 totalling 965ms (1.0% uptime), avg 965.0ms incr 3264 totalling 42,279ms (33.0% uptime), avg 13.0ms tenures 2,497 (avg 1 GCs/tenure) root table 0 overflows

1175: spur **Memory** old +0 bytes young +340,048 bytes used +340,048 bytes free -340,048 bytes **GCs** full 7 totalling 145,003ms (66.0% uptime), avg 20715.0ms incr 3288 totalling 30,912ms (14.0% uptime), avg 9.0ms tenures 7,146,505 (avg 0 GCs/tenure) root table 0 overflows

Total GC time 504: 43 secs 1175: 176 secs

See the performance reports attached.

I let VM people take care of the issue ;)

Cheers, Vincent

-----Original Message----- From: moose-dev-bounces@list.inf.unibe.ch [mailto:moose-dev-bounces@list.inf.unibe.ch] On Behalf Of Tudor Girba Sent: dimanche 24 janvier 2016 09:08 To: Moose-related development Subject: [Moose-dev] Re: mse loading looks slower :(

Hi,

I am talking about the difference between Moose 6 images: - October 7: https://ci.inria.fr/moose/job/moose-6.0/504/artifact/moose-6.0.zip

- yesterday: https://ci.inria.fr/moose/job/moose-6.0/1175/artifact/moose-6.0.zip

Multiple things did change, but not in Moose. In the end, I would like to understand where the slowness comes. Maybe it comes from Spur itself, but maybe it comes from somewhere else.

Cheers, Doru

...

On Jan 24, 2016, at 1:41 AM, Mariano Martinez Peck <marianopeck@gmail.com> wrote: Doru...just to be sure it is not a Pharo (image change), when you said before and after Spur, do you mean a Pharo 5.0 exactly (just before Spur) and a Pharo JUST after it? Otherwise, the slowness may come from the difference between the 2 Pharos you are running. Cheers,

On Sat, Jan 23, 2016 at 5:55 PM, Tudor Girba <tudor@tudorgirba.com> wrote: Hi,

I am doing some performance testing of Moose with the Spur VM on Mac.

I tried to load an MSE file with ArgoUML 0.34 and on my machine it loads twice as slow with Spur than before: - PreSpur: 0:00:01:07.272 - Spur: 0:00:02:10.508

Here is the reference file: https://dl.dropboxusercontent.com/u/18323746/Tmp/ArgoUML-0-34.mse.zip

And here is the script: [ MooseModel new importFromMSEStream: (StandardFileStream readOnlyFileNamed: (FileSystem workingDirectory / 'src' / 'ArgoUML-0-34' / 'ArgoUML-0-34.mse') fullName). ] timeToRun

Do you get the same?

Cheers, Doru

-- www.tudorgirba.com www.feenk.com

"Problem solving should be focused on describing the problem in a way that makes the solution obvious."

_______________________________________________ Moose-dev mailing list Moose-dev@list.inf.unibe.ch https://www.list.inf.unibe.ch/listinfo/moose-dev

-- Mariano http://marianopeck.wordpress.com _______________________________________________ Moose-dev mailing list Moose-dev@list.inf.unibe.ch https://www.list.inf.unibe.ch/listinfo/moose-dev -- www.tudorgirba.com www.feenk.com

"What is more important: To be happy, or to make happy?"

_______________________________________________ Moose-dev mailing list Moose-dev@list.inf.unibe.ch https://www.list.inf.unibe.ch/listinfo/moose-dev

On 24 Jan 2016, at 16:09, Stephan Eggermont <stephan@stack.nl> wrote:

On 24-01-16 15:34, Vincent BLONDEAU wrote:

...

Yes, and we have to do several passes to construct the model with references (so we store a mapping table). Maybe a collection or two are GCed.

At least a lot of Announcements and Announcers

Stephan

Of course it is good to optimise code (generating too much garbage is a delayed slowdown), but if GC is slower in Spur then that sounds like a regression that has to be studied. Sven

I think the use of SmallDictionaries for storing Famix state is not such a good idea. After loading the model: bag := Bag new . SmallDictionary allInstances do: [ :each | bag add: each size ]. bag sortedCounts "{530387->0. 19253->3. 18292->1. 20->2. 2->4. 2->11. 1->12}" bag := Bag new . SmallDictionary allInstances do: [ :each | bag add: (each keys) ]. bag sortedCounts an Array(530387->#() 19253->#(#cyclomaticComplexity #numberOfStatements #mooseName) 9311->#(#binding) 8707->#(#mooseName) 426->#(#comments) 20->#(#comments #ir) 13->#(#ir) 4->#(#comments #binding) 3->#(#binding #semanticWarning) 2->#('charset') 1->#(1 2 3 4 5 6 7 8 9 10 11 12) 1->#(14 15 16 17 18 19 20 21 22 23 24) 1->#(9 10 11 12 13 14 15 16) 1->an Array('-source-' #'`@method:' RBPatternVariableNode(`@args) '-pragmas-') 1->#(1 2 3 4 5 6 7 8 9 10 11) 1->{MooseModel} 1->{FAMIXAnnotationType. FAMIXClass. FAMIXNamespace. FAMIXMethod} 1->#(1 2))

Hi Eliot, Thanks for looking at this. Please let us know what support you need, or what kind of experiments we can do on our side. Cheers, Doru

On Jan 24, 2016, at 9:20 PM, Eliot Miranda <eliot.miranda@gmail.com> wrote:

Hi Vincent,

I'll take a look early this week. There's clearly a bug; the Spur GC is /not/ collecting those dictionaries :-( (thanks Stephan!). Assuming the bug is fixed times should come down (see below). It may be that the bug in Slang that I introduced in December has broken things because I don't see these symptoms in my daily work (but I use the most up-to-date VM version possible ;-) ). But I'm not in denial and look forward to using MOOSE as a good stress case.

I do want to say that the GC is not complete. Right now we have a scavenger that works well, and a global GC that has a slow compaction algorithm, and hence there are significant pauses. For example here's what I see as typical in using SPur for VMMaker work:

memory 160,432,128 bytes old 153,658,624 bytes (95.8%) young 4,838,224 bytes (3%) used 127,009,928 bytes (79.2%) free 28,126,456 bytes (17.5%) GCs 7,265 (?? ms between GCs) full 36 totalling 13,229 ms (0% uptime), avg 367.5 ms incr 7,229 totalling 6,546 ms (0% uptime), avg 0.9 ms tenures 3,589,063 (avg 0 GCs/tenure)

(There's no uptime in the above stats because we're still transitioning Squeak to the 64-bit clock and there are consequently bugs in computing uptime).

The plan is to add an incremental global GC so this work is broken up into much smaller pieces. I don't want to see 700ms pauses in global GC; one can't do game animation with that. So an incremental mark-sweep is needed. There are two nice papers we're considering, one from Lua and one for a truly concurrent collector. But time is pressing, so if anyone out there knows GC and is interested in helping this is a nicely self-contained project for which we'd love to have volunteers.

_,,,^..^,,,_ (phone)

...

On Jan 24, 2016, at 3:43 AM, Vincent BLONDEAU <vincent.blondeau@polytech-lille.net> wrote:

Hi,

I made the benchmarks with the files you provided. I have more or less the same magnitude: Version 504: 0:00:01:09.021 Version 1175: 0:00:02:37.507

However, by launching it in the time profiler (MooseModel new importFromMSEStream: (StandardFileStream readOnlyFileNamed: 'd:/ArgoUML-0-34.mse')), it takes 504: 1 min 55 1175: 4 min 25 Well there is a delta...

After investigation, the standard process has almost the same duration (120 secs for prespur and 140 secs for spur). But, there is a large difference in GC time:

504: not spur **Memory** old +144,822,000 bytes young -8,293,660 bytes used +136,528,340 bytes free -104,186,788 bytes

**GCs** full 1 totalling 965ms (1.0% uptime), avg 965.0ms incr 3264 totalling 42,279ms (33.0% uptime), avg 13.0ms tenures 2,497 (avg 1 GCs/tenure) root table 0 overflows

1175: spur **Memory** old +0 bytes young +340,048 bytes used +340,048 bytes free -340,048 bytes **GCs** full 7 totalling 145,003ms (66.0% uptime), avg 20715.0ms incr 3288 totalling 30,912ms (14.0% uptime), avg 9.0ms tenures 7,146,505 (avg 0 GCs/tenure) root table 0 overflows

Total GC time 504: 43 secs 1175: 176 secs

See the performance reports attached.

I let VM people take care of the issue ;)

Cheers, Vincent

-----Original Message----- From: moose-dev-bounces@list.inf.unibe.ch [mailto:moose-dev-bounces@list.inf.unibe.ch] On Behalf Of Tudor Girba Sent: dimanche 24 janvier 2016 09:08 To: Moose-related development Subject: [Moose-dev] Re: mse loading looks slower :(

Hi,

I am talking about the difference between Moose 6 images: - October 7: https://ci.inria.fr/moose/job/moose-6.0/504/artifact/moose-6.0.zip

- yesterday: https://ci.inria.fr/moose/job/moose-6.0/1175/artifact/moose-6.0.zip

Multiple things did change, but not in Moose. In the end, I would like to understand where the slowness comes. Maybe it comes from Spur itself, but maybe it comes from somewhere else.

Cheers, Doru

...

...

On Jan 24, 2016, at 1:41 AM, Mariano Martinez Peck <marianopeck@gmail.com> wrote:

Doru...just to be sure it is not a Pharo (image change), when you said before and after Spur, do you mean a Pharo 5.0 exactly (just before Spur) and a Pharo JUST after it? Otherwise, the slowness may come from the difference between the 2 Pharos you are running.

Cheers,

On Sat, Jan 23, 2016 at 5:55 PM, Tudor Girba <tudor@tudorgirba.com> wrote: Hi,

I am doing some performance testing of Moose with the Spur VM on Mac.

I tried to load an MSE file with ArgoUML 0.34 and on my machine it loads twice as slow with Spur than before: - PreSpur: 0:00:01:07.272 - Spur: 0:00:02:10.508

Here is the reference file: https://dl.dropboxusercontent.com/u/18323746/Tmp/ArgoUML-0-34.mse.zip

And here is the script: [ MooseModel new importFromMSEStream: (StandardFileStream readOnlyFileNamed: (FileSystem workingDirectory / 'src' / 'ArgoUML-0-34' / 'ArgoUML-0-34.mse') fullName). ] timeToRun

Do you get the same?

Cheers, Doru

-- www.tudorgirba.com www.feenk.com

"Problem solving should be focused on describing the problem in a way that makes the solution obvious."

_______________________________________________ Moose-dev mailing list Moose-dev@list.inf.unibe.ch https://www.list.inf.unibe.ch/listinfo/moose-dev

-- Mariano http://marianopeck.wordpress.com _______________________________________________ Moose-dev mailing list Moose-dev@list.inf.unibe.ch https://www.list.inf.unibe.ch/listinfo/moose-dev

-- www.tudorgirba.com www.feenk.com

"What is more important: To be happy, or to make happy?"

_______________________________________________ Moose-dev mailing list Moose-dev@list.inf.unibe.ch https://www.list.inf.unibe.ch/listinfo/moose-dev <report504.txt> <report1175.txt>

Hi Eliot, I created a package to easily reproduce the problem: https://dl.dropboxusercontent.com/u/18323746/Tmp/moose-6-for-GC-bug.zip In there you will get: - an image, - the MSE file to be loaded, and - a Playground script that loads that file. On my Mac, this takes some 3 minutes, while before it took about 1 minute. Let me know if you need anything else. Cheers, Doru

On Jan 24, 2016, at 9:33 PM, Eliot Miranda <eliot.miranda@gmail.com> wrote:

Hi Doru,

On Sun, Jan 24, 2016 at 12:26 PM, Tudor Girba <tudor@tudorgirba.com> wrote: Hi Eliot,

Thanks for looking at this.

Please let us know what support you need, or what kind of experiments we can do on our side.

What I want most of all is a test case that I can run on Mac. I'm assuming that I can just copy the job that Vincent mentioned in his email and use that as a test case. Do you see any issues with that? If so, how do I get a really bug MOOSE test case to run on Mac?

Cheers, Doru

...

On Jan 24, 2016, at 9:20 PM, Eliot Miranda <eliot.miranda@gmail.com> wrote:

Hi Vincent,

I'll take a look early this week. There's clearly a bug; the Spur GC is /not/ collecting those dictionaries :-( (thanks Stephan!). Assuming the bug is fixed times should come down (see below). It may be that the bug in Slang that I introduced in December has broken things because I don't see these symptoms in my daily work (but I use the most up-to-date VM version possible ;-) ). But I'm not in denial and look forward to using MOOSE as a good stress case.

I do want to say that the GC is not complete. Right now we have a scavenger that works well, and a global GC that has a slow compaction algorithm, and hence there are significant pauses. For example here's what I see as typical in using SPur for VMMaker work:

memory 160,432,128 bytes old 153,658,624 bytes (95.8%) young 4,838,224 bytes (3%) used 127,009,928 bytes (79.2%) free 28,126,456 bytes (17.5%) GCs 7,265 (?? ms between GCs) full 36 totalling 13,229 ms (0% uptime), avg 367.5 ms incr 7,229 totalling 6,546 ms (0% uptime), avg 0.9 ms tenures 3,589,063 (avg 0 GCs/tenure)

(There's no uptime in the above stats because we're still transitioning Squeak to the 64-bit clock and there are consequently bugs in computing uptime).

The plan is to add an incremental global GC so this work is broken up into much smaller pieces. I don't want to see 700ms pauses in global GC; one can't do game animation with that. So an incremental mark-sweep is needed. There are two nice papers we're considering, one from Lua and one for a truly concurrent collector. But time is pressing, so if anyone out there knows GC and is interested in helping this is a nicely self-contained project for which we'd love to have volunteers.

_,,,^..^,,,_ (phone)

...

On Jan 24, 2016, at 3:43 AM, Vincent BLONDEAU <vincent.blondeau@polytech-lille.net> wrote:

Hi,

I made the benchmarks with the files you provided. I have more or less the same magnitude: Version 504: 0:00:01:09.021 Version 1175: 0:00:02:37.507

However, by launching it in the time profiler (MooseModel new importFromMSEStream: (StandardFileStream readOnlyFileNamed: 'd:/ArgoUML-0-34.mse')), it takes 504: 1 min 55 1175: 4 min 25 Well there is a delta...

After investigation, the standard process has almost the same duration (120 secs for prespur and 140 secs for spur). But, there is a large difference in GC time:

504: not spur **Memory** old +144,822,000 bytes young -8,293,660 bytes used +136,528,340 bytes free -104,186,788 bytes

**GCs** full 1 totalling 965ms (1.0% uptime), avg 965.0ms incr 3264 totalling 42,279ms (33.0% uptime), avg 13.0ms tenures 2,497 (avg 1 GCs/tenure) root table 0 overflows

1175: spur **Memory** old +0 bytes young +340,048 bytes used +340,048 bytes free -340,048 bytes **GCs** full 7 totalling 145,003ms (66.0% uptime), avg 20715.0ms incr 3288 totalling 30,912ms (14.0% uptime), avg 9.0ms tenures 7,146,505 (avg 0 GCs/tenure) root table 0 overflows

Total GC time 504: 43 secs 1175: 176 secs

See the performance reports attached.

I let VM people take care of the issue ;)

Cheers, Vincent

-----Original Message----- From: moose-dev-bounces@list.inf.unibe.ch [mailto:moose-dev-bounces@list.inf.unibe.ch] On Behalf Of Tudor Girba Sent: dimanche 24 janvier 2016 09:08 To: Moose-related development Subject: [Moose-dev] Re: mse loading looks slower :(

Hi,

I am talking about the difference between Moose 6 images: - October 7: https://ci.inria.fr/moose/job/moose-6.0/504/artifact/moose-6.0.zip

- yesterday: https://ci.inria.fr/moose/job/moose-6.0/1175/artifact/moose-6.0.zip

Multiple things did change, but not in Moose. In the end, I would like to understand where the slowness comes. Maybe it comes from Spur itself, but maybe it comes from somewhere else.

Cheers, Doru

...

...

On Jan 24, 2016, at 1:41 AM, Mariano Martinez Peck <marianopeck@gmail.com> wrote:

Doru...just to be sure it is not a Pharo (image change), when you said before and after Spur, do you mean a Pharo 5.0 exactly (just before Spur) and a Pharo JUST after it? Otherwise, the slowness may come from the difference between the 2 Pharos you are running.

Cheers,

On Sat, Jan 23, 2016 at 5:55 PM, Tudor Girba <tudor@tudorgirba.com> wrote: Hi,

I am doing some performance testing of Moose with the Spur VM on Mac.

I tried to load an MSE file with ArgoUML 0.34 and on my machine it loads twice as slow with Spur than before: - PreSpur: 0:00:01:07.272 - Spur: 0:00:02:10.508

Here is the reference file: https://dl.dropboxusercontent.com/u/18323746/Tmp/ArgoUML-0-34.mse.zip

And here is the script: [ MooseModel new importFromMSEStream: (StandardFileStream readOnlyFileNamed: (FileSystem workingDirectory / 'src' / 'ArgoUML-0-34' / 'ArgoUML-0-34.mse') fullName). ] timeToRun

Do you get the same?

Cheers, Doru

-- www.tudorgirba.com www.feenk.com

"Problem solving should be focused on describing the problem in a way that makes the solution obvious."

_______________________________________________ Moose-dev mailing list Moose-dev@list.inf.unibe.ch https://www.list.inf.unibe.ch/listinfo/moose-dev

-- Mariano http://marianopeck.wordpress.com _______________________________________________ Moose-dev mailing list Moose-dev@list.inf.unibe.ch https://www.list.inf.unibe.ch/listinfo/moose-dev

-- www.tudorgirba.com www.feenk.com

"What is more important: To be happy, or to make happy?"

_______________________________________________ Moose-dev mailing list Moose-dev@list.inf.unibe.ch https://www.list.inf.unibe.ch/listinfo/moose-dev <report504.txt> <report1175.txt>

-- _,,,^..^,,,_ best, Eliot _______________________________________________ Moose-dev mailing list Moose-dev@list.inf.unibe.ch https://www.list.inf.unibe.ch/listinfo/moose-dev

Hi,

On Jan 24, 2016, at 9:51 PM, Eliot Miranda <eliot.miranda@gmail.com> wrote:

On Sun, Jan 24, 2016 at 12:42 PM, Tudor Girba <tudor@tudorgirba.com> wrote: Hi Eliot,

I created a package to easily reproduce the problem: https://dl.dropboxusercontent.com/u/18323746/Tmp/moose-6-for-GC-bug.zip

In there you will get: - an image, - the MSE file to be loaded, and - a Playground script that loads that file.

That sounds perfect.

On my Mac, this takes some 3 minutes, while before it took about 1 minute.

What's the system info on your Mac? Mine is "Version 10.9.5 Processor 2.2GHz Intel Core i7 Memory 16Gb 1333 MHz DDR3”

Model Name: MacBook Pro Model Identifier: MacBookPro10,1 Processor Name: Intel Core i7 Processor Speed: 2.6 GHz Number of Processors: 1 Total Number of Cores: 4 L2 Cache (per Core): 256 KB L3 Cache: 6 MB Memory: 16 GB Boot ROM Version: MBP101.00EE.B0A SMC Version (system): 2.3f36 Serial Number (system): C02JG3T5DKQ5 Hardware UUID: 1FF116A4-5040-53B2-A133-F95256DE6C94 Cheers, Doru

Let me know if you need anything else.

Cheers, Doru

...

On Jan 24, 2016, at 9:33 PM, Eliot Miranda <eliot.miranda@gmail.com> wrote:

Hi Doru,

On Sun, Jan 24, 2016 at 12:26 PM, Tudor Girba <tudor@tudorgirba.com> wrote: Hi Eliot,

Thanks for looking at this.

Please let us know what support you need, or what kind of experiments we can do on our side.

What I want most of all is a test case that I can run on Mac. I'm assuming that I can just copy the job that Vincent mentioned in his email and use that as a test case. Do you see any issues with that? If so, how do I get a really bug MOOSE test case to run on Mac?

Cheers, Doru

...

On Jan 24, 2016, at 9:20 PM, Eliot Miranda <eliot.miranda@gmail.com> wrote:

Hi Vincent,

I'll take a look early this week. There's clearly a bug; the Spur GC is /not/ collecting those dictionaries :-( (thanks Stephan!). Assuming the bug is fixed times should come down (see below). It may be that the bug in Slang that I introduced in December has broken things because I don't see these symptoms in my daily work (but I use the most up-to-date VM version possible ;-) ). But I'm not in denial and look forward to using MOOSE as a good stress case.

I do want to say that the GC is not complete. Right now we have a scavenger that works well, and a global GC that has a slow compaction algorithm, and hence there are significant pauses. For example here's what I see as typical in using SPur for VMMaker work:

memory 160,432,128 bytes old 153,658,624 bytes (95.8%) young 4,838,224 bytes (3%) used 127,009,928 bytes (79.2%) free 28,126,456 bytes (17.5%) GCs 7,265 (?? ms between GCs) full 36 totalling 13,229 ms (0% uptime), avg 367.5 ms incr 7,229 totalling 6,546 ms (0% uptime), avg 0.9 ms tenures 3,589,063 (avg 0 GCs/tenure)

(There's no uptime in the above stats because we're still transitioning Squeak to the 64-bit clock and there are consequently bugs in computing uptime).

The plan is to add an incremental global GC so this work is broken up into much smaller pieces. I don't want to see 700ms pauses in global GC; one can't do game animation with that. So an incremental mark-sweep is needed. There are two nice papers we're considering, one from Lua and one for a truly concurrent collector. But time is pressing, so if anyone out there knows GC and is interested in helping this is a nicely self-contained project for which we'd love to have volunteers.

_,,,^..^,,,_ (phone)

...

On Jan 24, 2016, at 3:43 AM, Vincent BLONDEAU <vincent.blondeau@polytech-lille.net> wrote:

Hi,

I made the benchmarks with the files you provided. I have more or less the same magnitude: Version 504: 0:00:01:09.021 Version 1175: 0:00:02:37.507

However, by launching it in the time profiler (MooseModel new importFromMSEStream: (StandardFileStream readOnlyFileNamed: 'd:/ArgoUML-0-34.mse')), it takes 504: 1 min 55 1175: 4 min 25 Well there is a delta...

After investigation, the standard process has almost the same duration (120 secs for prespur and 140 secs for spur). But, there is a large difference in GC time:

504: not spur **Memory** old +144,822,000 bytes young -8,293,660 bytes used +136,528,340 bytes free -104,186,788 bytes

**GCs** full 1 totalling 965ms (1.0% uptime), avg 965.0ms incr 3264 totalling 42,279ms (33.0% uptime), avg 13.0ms tenures 2,497 (avg 1 GCs/tenure) root table 0 overflows

1175: spur **Memory** old +0 bytes young +340,048 bytes used +340,048 bytes free -340,048 bytes **GCs** full 7 totalling 145,003ms (66.0% uptime), avg 20715.0ms incr 3288 totalling 30,912ms (14.0% uptime), avg 9.0ms tenures 7,146,505 (avg 0 GCs/tenure) root table 0 overflows

Total GC time 504: 43 secs 1175: 176 secs

See the performance reports attached.

I let VM people take care of the issue ;)

Cheers, Vincent

-----Original Message----- From: moose-dev-bounces@list.inf.unibe.ch [mailto:moose-dev-bounces@list.inf.unibe.ch] On Behalf Of Tudor Girba Sent: dimanche 24 janvier 2016 09:08 To: Moose-related development Subject: [Moose-dev] Re: mse loading looks slower :(

Hi,

I am talking about the difference between Moose 6 images: - October 7: https://ci.inria.fr/moose/job/moose-6.0/504/artifact/moose-6.0.zip

- yesterday: https://ci.inria.fr/moose/job/moose-6.0/1175/artifact/moose-6.0.zip

Multiple things did change, but not in Moose. In the end, I would like to understand where the slowness comes. Maybe it comes from Spur itself, but maybe it comes from somewhere else.

Cheers, Doru

...

...

On Jan 24, 2016, at 1:41 AM, Mariano Martinez Peck <marianopeck@gmail.com> wrote:

Doru...just to be sure it is not a Pharo (image change), when you said before and after Spur, do you mean a Pharo 5.0 exactly (just before Spur) and a Pharo JUST after it? Otherwise, the slowness may come from the difference between the 2 Pharos you are running.

Cheers,

On Sat, Jan 23, 2016 at 5:55 PM, Tudor Girba <tudor@tudorgirba.com> wrote: Hi,

I am doing some performance testing of Moose with the Spur VM on Mac.

I tried to load an MSE file with ArgoUML 0.34 and on my machine it loads twice as slow with Spur than before: - PreSpur: 0:00:01:07.272 - Spur: 0:00:02:10.508

Here is the reference file: https://dl.dropboxusercontent.com/u/18323746/Tmp/ArgoUML-0-34.mse.zip

And here is the script: [ MooseModel new importFromMSEStream: (StandardFileStream readOnlyFileNamed: (FileSystem workingDirectory / 'src' / 'ArgoUML-0-34' / 'ArgoUML-0-34.mse') fullName). ] timeToRun

Do you get the same?

Cheers, Doru

-- www.tudorgirba.com www.feenk.com

"Problem solving should be focused on describing the problem in a way that makes the solution obvious."

_______________________________________________ Moose-dev mailing list Moose-dev@list.inf.unibe.ch https://www.list.inf.unibe.ch/listinfo/moose-dev

-- Mariano http://marianopeck.wordpress.com _______________________________________________ Moose-dev mailing list Moose-dev@list.inf.unibe.ch https://www.list.inf.unibe.ch/listinfo/moose-dev

-- www.tudorgirba.com www.feenk.com

"What is more important: To be happy, or to make happy?"

_______________________________________________ Moose-dev mailing list Moose-dev@list.inf.unibe.ch https://www.list.inf.unibe.ch/listinfo/moose-dev <report504.txt> <report1175.txt>

-- _,,,^..^,,,_ best, Eliot _______________________________________________ Moose-dev mailing list Moose-dev@list.inf.unibe.ch https://www.list.inf.unibe.ch/listinfo/moose-dev

-- _,,,^..^,,,_ best, Eliot _______________________________________________ Moose-dev mailing list Moose-dev@list.inf.unibe.ch https://www.list.inf.unibe.ch/listinfo/moose-dev

Hi Eliot, Indeed, this was a Spur image. I will put together the same setup with a pre-Spur image and get back to you. Cheers, Doru

On Jan 24, 2016, at 11:35 PM, Eliot Miranda <eliot.miranda@gmail.com> wrote:

Hi Doru,

On Sun, Jan 24, 2016 at 12:42 PM, Tudor Girba <tudor@tudorgirba.com> wrote: Hi Eliot,

I created a package to easily reproduce the problem: https://dl.dropboxusercontent.com/u/18323746/Tmp/moose-6-for-GC-bug.zip

In there you will get: - an image, - the MSE file to be loaded, and - a Playground script that loads that file.

I take it that this is a Spur image. t would be great to have both a Spur and a V3 mage to compare, as similar as is easily achievable. But as long as there's both a V3 and a Spur image...

On my Mac, this takes some 3 minutes, while before it took about 1 minute.

Let me know if you need anything else.

Cheers, Doru

...

On Jan 24, 2016, at 9:33 PM, Eliot Miranda <eliot.miranda@gmail.com> wrote:

Hi Doru,

On Sun, Jan 24, 2016 at 12:26 PM, Tudor Girba <tudor@tudorgirba.com> wrote: Hi Eliot,

Thanks for looking at this.

Please let us know what support you need, or what kind of experiments we can do on our side.

What I want most of all is a test case that I can run on Mac. I'm assuming that I can just copy the job that Vincent mentioned in his email and use that as a test case. Do you see any issues with that? If so, how do I get a really bug MOOSE test case to run on Mac?

Cheers, Doru

...

On Jan 24, 2016, at 9:20 PM, Eliot Miranda <eliot.miranda@gmail.com> wrote:

Hi Vincent,

I'll take a look early this week. There's clearly a bug; the Spur GC is /not/ collecting those dictionaries :-( (thanks Stephan!). Assuming the bug is fixed times should come down (see below). It may be that the bug in Slang that I introduced in December has broken things because I don't see these symptoms in my daily work (but I use the most up-to-date VM version possible ;-) ). But I'm not in denial and look forward to using MOOSE as a good stress case.

I do want to say that the GC is not complete. Right now we have a scavenger that works well, and a global GC that has a slow compaction algorithm, and hence there are significant pauses. For example here's what I see as typical in using SPur for VMMaker work:

memory 160,432,128 bytes old 153,658,624 bytes (95.8%) young 4,838,224 bytes (3%) used 127,009,928 bytes (79.2%) free 28,126,456 bytes (17.5%) GCs 7,265 (?? ms between GCs) full 36 totalling 13,229 ms (0% uptime), avg 367.5 ms incr 7,229 totalling 6,546 ms (0% uptime), avg 0.9 ms tenures 3,589,063 (avg 0 GCs/tenure)

(There's no uptime in the above stats because we're still transitioning Squeak to the 64-bit clock and there are consequently bugs in computing uptime).

The plan is to add an incremental global GC so this work is broken up into much smaller pieces. I don't want to see 700ms pauses in global GC; one can't do game animation with that. So an incremental mark-sweep is needed. There are two nice papers we're considering, one from Lua and one for a truly concurrent collector. But time is pressing, so if anyone out there knows GC and is interested in helping this is a nicely self-contained project for which we'd love to have volunteers.

_,,,^..^,,,_ (phone)

...

On Jan 24, 2016, at 3:43 AM, Vincent BLONDEAU <vincent.blondeau@polytech-lille.net> wrote:

Hi,

I made the benchmarks with the files you provided. I have more or less the same magnitude: Version 504: 0:00:01:09.021 Version 1175: 0:00:02:37.507

However, by launching it in the time profiler (MooseModel new importFromMSEStream: (StandardFileStream readOnlyFileNamed: 'd:/ArgoUML-0-34.mse')), it takes 504: 1 min 55 1175: 4 min 25 Well there is a delta...

After investigation, the standard process has almost the same duration (120 secs for prespur and 140 secs for spur). But, there is a large difference in GC time:

504: not spur **Memory** old +144,822,000 bytes young -8,293,660 bytes used +136,528,340 bytes free -104,186,788 bytes

**GCs** full 1 totalling 965ms (1.0% uptime), avg 965.0ms incr 3264 totalling 42,279ms (33.0% uptime), avg 13.0ms tenures 2,497 (avg 1 GCs/tenure) root table 0 overflows

1175: spur **Memory** old +0 bytes young +340,048 bytes used +340,048 bytes free -340,048 bytes **GCs** full 7 totalling 145,003ms (66.0% uptime), avg 20715.0ms incr 3288 totalling 30,912ms (14.0% uptime), avg 9.0ms tenures 7,146,505 (avg 0 GCs/tenure) root table 0 overflows

Total GC time 504: 43 secs 1175: 176 secs

See the performance reports attached.

I let VM people take care of the issue ;)

Cheers, Vincent

-----Original Message----- From: moose-dev-bounces@list.inf.unibe.ch [mailto:moose-dev-bounces@list.inf.unibe.ch] On Behalf Of Tudor Girba Sent: dimanche 24 janvier 2016 09:08 To: Moose-related development Subject: [Moose-dev] Re: mse loading looks slower :(

Hi,

I am talking about the difference between Moose 6 images: - October 7: https://ci.inria.fr/moose/job/moose-6.0/504/artifact/moose-6.0.zip

- yesterday: https://ci.inria.fr/moose/job/moose-6.0/1175/artifact/moose-6.0.zip

Multiple things did change, but not in Moose. In the end, I would like to understand where the slowness comes. Maybe it comes from Spur itself, but maybe it comes from somewhere else.

Cheers, Doru

...

...

On Jan 24, 2016, at 1:41 AM, Mariano Martinez Peck <marianopeck@gmail.com> wrote:

Doru...just to be sure it is not a Pharo (image change), when you said before and after Spur, do you mean a Pharo 5.0 exactly (just before Spur) and a Pharo JUST after it? Otherwise, the slowness may come from the difference between the 2 Pharos you are running.

Cheers,

On Sat, Jan 23, 2016 at 5:55 PM, Tudor Girba <tudor@tudorgirba.com> wrote: Hi,

I am doing some performance testing of Moose with the Spur VM on Mac.

I tried to load an MSE file with ArgoUML 0.34 and on my machine it loads twice as slow with Spur than before: - PreSpur: 0:00:01:07.272 - Spur: 0:00:02:10.508

Here is the reference file: https://dl.dropboxusercontent.com/u/18323746/Tmp/ArgoUML-0-34.mse.zip

And here is the script: [ MooseModel new importFromMSEStream: (StandardFileStream readOnlyFileNamed: (FileSystem workingDirectory / 'src' / 'ArgoUML-0-34' / 'ArgoUML-0-34.mse') fullName). ] timeToRun

Do you get the same?

Cheers, Doru

-- www.tudorgirba.com www.feenk.com

"Problem solving should be focused on describing the problem in a way that makes the solution obvious."

_______________________________________________ Moose-dev mailing list Moose-dev@list.inf.unibe.ch https://www.list.inf.unibe.ch/listinfo/moose-dev

-- Mariano http://marianopeck.wordpress.com _______________________________________________ Moose-dev mailing list Moose-dev@list.inf.unibe.ch https://www.list.inf.unibe.ch/listinfo/moose-dev

-- www.tudorgirba.com www.feenk.com

"What is more important: To be happy, or to make happy?"

_______________________________________________ Moose-dev mailing list Moose-dev@list.inf.unibe.ch https://www.list.inf.unibe.ch/listinfo/moose-dev <report504.txt> <report1175.txt>

-- _,,,^..^,,,_ best, Eliot _______________________________________________ Moose-dev mailing list Moose-dev@list.inf.unibe.ch https://www.list.inf.unibe.ch/listinfo/moose-dev

-- _,,,^..^,,,_ best, Eliot _______________________________________________ Moose-dev mailing list Moose-dev@list.inf.unibe.ch https://www.list.inf.unibe.ch/listinfo/moose-dev

Hi Stephan, On Sun, Jan 24, 2016 at 1:04 PM, Stephan Eggermont <stephan@stack.nl> wrote:

On 24-01-16 21:20, Eliot Miranda wrote:

...

Hi Vincent,

I'll take a look early this week. There's clearly a bug; the Spur GC is /not/ collecting those dictionaries :-( (thanks Stephan!).

I'm afraid the dictionaries are part of the model and should not be garbage collected. They do make things slow though, so we need a better model. Garbage collection cycle takes a few seconds now on my 4 GHz 4790K 32GB.

So Vincent's awful sats are no longer showing up? He reported a 140 second GC pause. What are you seeing? (.i.e. can you post the full stats?) If this has disappeared then I suspect my Slang bug, which I fixed yesterday. But I will investigate.

Stephan

_,,,^..^,,,_ best, Eliot