January 2016 - Pharo-dev - lists.pharo.org

Re: [Pharo-dev] Monitor>>critical: implementation

by Denis Kudriashov

2016-01-05 12:59 GMT+01:00 Max Leske <maxleske(a)gmail.com>: > You may have a point there. Although the monitor uses a Semaphore > #forMutualExclusion, which is a semaphore with one signal. So the case > where the semaphore gets signalled without #wait is less likely I think. > No. It has no relation to incorrect scenarios of semaphore signalling. #forMutualExclusion just make semaphore free to use. So first caller of #wait will grab lock. And others will wait. Two times signalled semaphore will allow two critical sections be performed simultaneously which should not happen

Jan. 5, 2016

Re: [Pharo-dev] Monitor>>critical: implementation

by Max Leske

> On 05 Jan 2016, at 11:17, Denis Kudriashov <dionisiydk(a)gmail.com> wrote: > > Hello. > > Here is Monitor code: > > critical: aBlock > ^[ > self enter. > aBlock value] > ensure: [self exit]. > > Now look at Semaphore comment and implementation: > > We need to catch eventual interruptions very carefully. > The naive approach of just doing, e.g.,: > self wait. > aBlock ensure: [self signal]. > will fail if the active process gets terminated while in the wait. > However, the equally naive: > [self wait. > aBlock value] ensure: [self signal]. > will fail too, since the active process may get interrupted while > entering the ensured block and leave the semaphore signaled twice. > > critical: aBlock > | blockValue caught | > caught := false. > [ > caught := true. > self wait. > blockValue := mutuallyExcludedBlock value > ] ensure: [caught ifTrue: [self signal]]. > ^blockValue > > So Monitor can be corrupted when process which enters critical section will be terminated. "Ensure block" can signal mutex which was not in waiting mode. Which means that next critical call in another Process will be not guarded. > > What you think? Is my investigation right? > > I look at users of Monitor. Most time it is used like Mutex. Only #critical: method is used. So Monitor can be replaced by Mutex at that places. And possible problem will be fixed. > > SharedQueue is classic example for Monitor usage. Interesting is anybody got problems with it? You may have a point there. Although the monitor uses a Semaphore #forMutualExclusion, which is a semaphore with one signal. So the case where the semaphore gets signalled without #wait is less likely I think. What I donât quite understand is why Monitor doesnât use Semaphore>>critical:. If it did that, then there wouldnât be a problem. The problem only exists because #wait and #signal are explicitly sent by the monitor.

Jan. 5, 2016

small refactor on files.pharo.org

by Esteban Lorenzano

I move old VMs from files.pharo.org/vm <http://files.pharo.org/vm> to files.pharo.org/vm/old <http://files.pharo.org/vm/old> because it was confusing and some students were lost (Stef too :P). cheers, Esteban

Jan. 5, 2016

Re: [Pharo-dev] [Ann] ReadWriteLock

by Denis Kudriashov

2016-01-04 18:51 GMT+01:00 Tudor Girba <tudor(a)tudorgirba.com>: > Is this meant to be part of the main distribution? I don't know. People should decide it. To me it is normal to get explicit dependency in configuration.

Jan. 5, 2016

Monitor>>critical: implementation

by Denis Kudriashov

Hello. Here is Monitor code: critical: aBlock > ^[ > self enter. > aBlock value] > ensure: [self exit]. Now look at Semaphore comment and implementation: > We need to catch eventual interruptions very carefully. > The naive approach of just doing, e.g.,: > self wait. > aBlock ensure: [self signal]. > will fail if the active process gets terminated while in the wait. However, the equally naive: > [self wait. > aBlock value] ensure: [self signal]. > will fail too, since the active process may get interrupted while > entering the ensured block and leave the semaphore signaled twice. > > critical: aBlock > | blockValue caught | > caught := false. > [ > caught := true. > self wait. > blockValue := mutuallyExcludedBlock value > ] ensure: [caught ifTrue: [self signal]]. > ^blockValue So Monitor can be corrupted when process which enters critical section will be terminated. "Ensure block" can signal mutex which was not in waiting mode. Which means that next critical call in another Process will be not guarded. What you think? Is my investigation right? I look at users of Monitor. Most time it is used like Mutex. Only #critical: method is used. So Monitor can be replaced by Mutex at that places. And possible problem will be fixed. SharedQueue is classic example for Monitor usage. Interesting is anybody got problems with it?

Jan. 5, 2016

Re: [Pharo-dev] [Ann] ReadWriteLock

by Denis Kudriashov

Hi. 2016-01-04 19:19 GMT+01:00 Igor Stasenko <siguctua(a)gmail.com>: > It heavily depends on data structure you using what kind of > synchronization mechanism is more efficient than another. > For instance AtomicQueue implementation is one that falls in such category > 'many readers, but only single writer'. But it's not using semaphores to > synchronise access to it. > Can you explain deeply what problem with AtomicQueue in 'many readers, but only single writer'?

Jan. 5, 2016

Reverted changes in Monticello ?

by Thierry Goubier

Hi all, I had an error with the metadata-less mode of gitfiletree, and I discovered that a fix[1], integrated in 50470, was reverted sometime afterwards. And when looking a bit more, 50510 has MonticelloFileTree-Core-TheIntegrator.134 while the latest in main is MonticelloFileTree-Core-TheIntegrator.193 (with not a few fixes). Anybody knows what is happening? Thierry [1] https://pharo.fogbugz.com/f/cases/17125/Metadata-less-compatibility-with-Fi…

Jan. 5, 2016

Re: [Pharo-dev] [Ann] ReadWriteLock

by Denis Kudriashov

Hi 2016-01-04 18:55 GMT+01:00 Max Leske <maxleske(a)gmail.com>: > Hi Denis, > > How does your ReadWriteLock compare to Monitor? I was under the impression > that what you describe could be done with a Monitor. Monitor is different synchronisation mechanism. It can be used to implement ReadWriteLock. But It looks like more expensive solution. After look at Pharo implementation of Monitor I realise that it #critical: method not takes into account issue described in Semaphore>>critical:. I will ask about it on separate thread.

Jan. 5, 2016

Re: [Pharo-dev] [Vm-dev] Repeatable VM crash

by Esteban Lorenzano

> On 04 Jan 2016, at 21:22, Mariano Martinez Peck <marianopeck(a)gmail.com> wrote: > > > > On Sun, Dec 7, 2014 at 3:27 PM, Eliot Miranda <eliot.miranda(a)gmail.com <mailto:eliot.miranda@gmail.com>> wrote: > Hi Mariano, > > On Dec 7, 2014, at 10:21 AM, Mariano Martinez Peck <marianopeck(a)gmail.com <mailto:marianopeck@gmail.com>> wrote: > >> >> >> On Sun, Dec 7, 2014 at 3:09 PM, Eliot Miranda <eliot.miranda(a)gmail.com <mailto:eliot.miranda@gmail.com>> wrote: >> Hi Hilaire, >> >> >> On Dec 7, 2014, at 12:35 AM, Hilaire <hilaire(a)drgeo.eu <mailto:hilaire@drgeo.eu>> wrote: >> >> > I download this one >> > http://www.mirandabanda.org/files/Cog/VM/VM.r3164/coglinux-14.48.3164.tgz <http://www.mirandabanda.org/files/Cog/VM/VM.r3164/coglinux-14.48.3164.tgz> >> > >> > and the crash is still there. Attached crash dump. >> >> different bugs. The 3164 VMs fix a 32-bit Spur bug. Your image is showing a bug in Cog that looks to be to do with the decompression primitive. I haven't had time to look at it deeply yet but can see it and should be able to fix it soon. >> >> You mean GZip stuff? Because I have found a similar issue years ago: http://forum.world.st/Re-Pharo-project-GZipWriteStream-crashing-my-VM-when-… <http://forum.world.st/Re-Pharo-project-GZipWriteStream-crashing-my-VM-when-…> > No. That bug got fixed recently. This is a bug in the custom huffman-ciding bitmap decompression routine (selector begins with "decompress") that's compiled to a primitive. > > > > Hi, > > It seems Stephan is seeing this bug again. > Eliot, is there a chance that the fix was not finally integrated or something? vm build is at least 1 month old (spur one). Cog vm is at least 6 months old (yet some stuff has been integrated into it, but not much). vm development is moving to spur branch so changes will be there (and not old cog branch) but to resume vm build we need to fix a bug in opal compiler (in spur context): https://pharo.fogbugz.com/f/cases/17251 <https://pharo.fogbugz.com/f/cases/17251> Soâ¦ probably, that bug fix is not present yet in the vm (I cannot be sure, but probably) Esteban > > > >> >> >> > >> > Thanks >> > >> > Hilaire >> > >> > Le 06/12/2014 22:57, Eliot Miranda a Ã©crit : >> >> Hi Esteban, >> >> >> >> plausibly the fix to the latest VM might fix your crash. You might >> >> test? Fingers crossed. >> >> >> >> On Fri, Dec 5, 2014 at 11:26 PM, Esteban Lorenzano <estebanlm(a)gmail.com <mailto:estebanlm@gmail.com> >> >> <mailto:estebanlm@gmail.com <mailto:estebanlm@gmail.com>>> wrote: >> >> >> >> yes, but the first step is to be sure you have the latest vm version :) >> >> >> >> Esteban >> > >> > >> > -- >> > Dr. Geo - http://drgeo.eu <http://drgeo.eu/> >> > iStoa - http://istoa.drgeo.eu <http://istoa.drgeo.eu/> >> > <crash.dmp> >> >> >> >> >> -- >> Mariano >> http://marianopeck.wordpress.com <http://marianopeck.wordpress.com/> > > > > -- > Mariano > http://marianopeck.wordpress.com <http://marianopeck.wordpress.com/>

Jan. 5, 2016

Re: [Pharo-dev] Some test where Spur more slow than Cog

by ClÃ©ment Bera

I could reproduce with Squeak 5 and Squeak 4.6 so I was able to use the VM profiler. I used: | prototype | prototype := BorderedMorph new. 10000 timesRepeat: [prototype veryDeepCopy] The profiler ran for ~0.8 second. Maybe you can have more accurate result with longer bench but it takes already a while to display results. *In Spur:* 1044 samples in the VM (1057 samples in the entire program) 98.77% of total 754 samples in generated vm code 72.22% of entire vm (71.33% of total) 290 samples in vanilla vm code 27.78% of entire vm (27.44% of total) % of generated vm code (% of total) (samples) (cumulative) 37.67% (26.87%) Dictionary>>associationsDo: (284) (37.67%) 30.37% (21.67%) Object>>at: (229) (68.04%) 3.18% ( 2.27%) WeakKeyDictionary>>associationsDo: (24) (71.22%) 2.12% ( 1.51%) MethodDictionary>>scanFor: (16) (73.34%) 1.59% ( 1.14%) SmallInteger>>\\ (12) (74.93%) % of vanilla vm code (% of total) (samples) (cumulative) 58.62% (16.08%) scavengeReferentsOf (170) (58.62%) 24.48% ( 6.72%) instantiateClassindexableSize (71) (83.10%) 2.76% ( 0.76%) copyAndForward (8) (85.86%) 2.07% ( 0.57%) doScavenge (6) (87.93%) 2.07% ( 0.57%) lookupOrdinaryNoMNUEtcInClass (6) (90.00%) 13 samples in the rest 1.23% of total % of rest (% of total) (samples) (cumulative) 53.85% ( 0.66%) SCDynamicStoreCreate (7) (53.85%) 46.15% ( 0.57%) _platform_memmove$VARIANT$sse42 (6) (100.0%) 46.15% ( 0.57%) OSAtomicAdd32Barrier (6) (146.15%) **GCs** full 0 totalling 0ms (0% elapsed time) scavenges 258 totalling 151ms (18.97% elapsed time), avg 0.585ms *In SqueakV3:* 1189 samples in the VM (1196 samples in the entire program) 99.41% of total 998 samples in generated vm code 83.94% of entire vm (83.44% of total) 191 samples in vanilla vm code 16.06% of entire vm (15.97% of total) % of generated vm code (% of total) (samples) (cumulative) 43.19% (36.04%) Object>>at: (431) (43.19%) 35.07% (29.26%) Dictionary>>associationsDo: (350) (78.26%) 1.90% ( 1.59%) WeakKeyDictionary>>associationsDo: (19) (80.16%) 1.30% ( 1.09%) Object>>veryDeepCopyWith: (13) (81.46%) 1.20% ( 1.00%) BlockClosure>>value (12) (82.67%) % of vanilla vm code (% of total) (samples) (cumulative) 30.37% ( 4.85%) instantiateClassindexableSize (58) (30.37%) 13.61% ( 2.17%) incrementalGC (26) (43.98%) 9.42% ( 1.51%) markAndTrace (18) (53.40%) 8.38% ( 1.34%) marryFrameSP (16) (61.78%) 6.81% ( 1.09%) eeInstantiateSmallClassnumSlots (13) (68.59%) 5.24% ( 0.84%) createClosureNumArgsnumCopiedstartpc (10) (73.82%) 4.71% ( 0.75%) mapPointersInObjectsFromto (9) (78.53%) **GCs** full 0 totalling 0ms (0% elapsed time) incr 270 totalling 46ms (5.094% elapsed time), avg 0.17ms *Conclusion ?* It feels that the main issue is with the scavenges that take overall more time in Spur in this case. Especially, the method scavengeReferentsOf seems to be guilty and may be worth looking at. There are other details, like now a little bit of execution time seems to be used to allocate new memory segments and associationsDo: looks slower, likely because of == and forwarders, but this is not as important. What do you think Eliot ? 2015-12-19 11:06 GMT+01:00 Denis Kudriashov <dionisiydk(a)gmail.com>: > Hi > > 2015-12-19 4:18 GMT+01:00 Eliot Miranda <eliot.miranda(a)gmail.com>: > >> and used in some copying code to access indexable fields. Chris Muller >> found this in some of his Magma code and so we added a pair of new >> primitives, primitiveSlotAt and primitiveSlotAtPut that access any slot >> within the sequence of named followed by indexable inst vars, avoiding the >> primitive failure. So in Squeak Spur instVarAt: now reads: >> >> instVarAt: index >> "Primitive. Answer a fixed variable in an object. The numbering of the >> variables >> corresponds to the named instance variables, followed by the indexed >> instance >> variables. Fail if the index is not an Integer or is not the index of a >> fixed variable. >> Essential. See Object documentation whatIsAPrimitive." >> >> <primitive: 173 error: ec> >> self primitiveFailed >> > > Same code used in Pharo too. > > I try same test on latest squeak images: > > Cog: '10,100 per second. 98.8 microseconds per run.' > Spur: '7,610 per second. 131 microseconds per run.' >

Jan. 5, 2016