- Pharo-dev - lists.pharo.org

[Pharo 7.0-dev] Build #975: 21959-Package-Slot-Traits-is-empty-and-can-be-removed

by ci-pharo-ci-jenkins2＠inria.fr

There is a new Pharo build available! The status of the build #975 was: FAILURE. The Pull Request #1414 was integrated: "21959-Package-Slot-Traits-is-empty-and-can-be-removed" Pull request url: https://github.com/pharo-project/pharo/pull/1414 Issue Url: https://pharo.fogbugz.com/f/cases/21959 Build Url: https://ci.inria.fr/pharo-ci-jenkins2/job/Test%20pending%20pull%20request%2…

May 28, 2018

[Pharo 7.0-dev] Build #974: 22006-add-startUsingOverlayForDevelopment-stopUsingOverlayForDevelopment

by ci-pharo-ci-jenkins2＠inria.fr

There is a new Pharo build available! The status of the build #974 was: SUCCESS. The Pull Request #1449 was integrated: "22006-add-startUsingOverlayForDevelopment-stopUsingOverlayForDevelopment" Pull request url: https://github.com/pharo-project/pharo/pull/1449 Issue Url: https://pharo.fogbugz.com/f/cases/22006 Build Url: https://ci.inria.fr/pharo-ci-jenkins2/job/Test%20pending%20pull%20request%2…

May 28, 2018

Re: [Pharo-dev] Curious context change latency experiment results (wrt DelayScheduler)

by Eliot Miranda

Hi Ben, _,,,^..^,,,_ (phone) > On May 26, 2018, at 10:46 PM, Ben Coman <btc(a)openinworld.com> wrote: > > > >> On 6 May 2018 at 04:20, Eliot Miranda <eliot.miranda(a)gmail.com> wrote: >> Hi Ben, >> >> >>> On May 5, 2018, at 7:25 AM, Ben Coman <btc(a)openinworld.com> wrote: >>> >>> >>> >>>> On 5 May 2018 at 22:10, Ben Coman <btc(a)openinworld.com> wrote: >>>> One of the key parts of Delay scheduling is setting the resumption time. >>>> There are two places this could be done. >>>> a. In Delay>>#schedule, running (for example) at userSchedulingPriority=40 >>>> b. In DelayMicrosecondScheduler>>#handleTImerEvent:, running at timingPriority=80 >>>> >>>> When we were using the millisecond clock for delay scheduling, >>>> it made sense to use (b.) since the clock might(?) roll over >>>> between when resumption time was set, and when #handleTImerEvent: expires delays. >>>> >>>> This should not be a problem now that we are using the microsecond clock, >>>> so I wondered what the latency penalty might be between (a.) and (b.). >>>> I ran a little experiment that indicates the cost/latency of switching threads, >>>> and was curious if anyone can comment on the validity of the experiment >>>> and interpretation of results. >>>> >>>> I made the following three changes... >>>> >>>> DelayScheduler subclass: #DelayMicrosecondScheduler >>>> instanceVariableNames: 'latencyStart countSinceLastZero' >>>> classVariableNames: 'LatencyCounts' >>>> poolDictionaries: '' >>>> category: 'Kernel-Processes' >>>> >>>> >>>> DelayMicrosecondScheduler>>#schedule: aDelay >>>> latencyStart:= Time primUTCMicrosecondsClock. "This is position (a.)" >>>> aDelay schedulerBeingWaitedOn ifTrue: [^self error: 'This Delay has already been scheduled.']. >>>> accessProtect critical: [ >>>> scheduledDelay := aDelay. >>>> timingSemaphore signal. "#handleTimerEvent: sets scheduledDelay:=nil" >>>> ]. >>>> >>>> >>>> DelayMicrosecondScheduler>>#handleTimerEvent: microsecondNowTick >>>> | microsecondNextTick | >>>> "Process any schedule requests" "This is position (b.)" >>>> scheduledDelay ifNotNil: [ >>>> |latency| >>>> latency := Time primUTCMicrosecondsClock - latencyStart. >>>> LatencyCounts ifNil: [ LatencyCounts := Bag new ]. >>>> LatencyCounts add: latency. >>>> latency = 0 >>>> ifTrue: [ countSinceLastZero := 1 + (countSinceLastZero ifNil: [0])] >>>> ifFalse: [Transcript >>>> crShow: 'zero latency count ' , countSinceLastZero printString ; >>>> show: ' before latency ', latency printString. >>>> countSinceLastZero := 0]. >>>> "Schedule the given delay." >>>> scheduledDelay scheduler: self resumptionTime: microsecondNowTick + (1000 * scheduledDelay millisecondDelayDuration). >>>> self scheduleDelay: scheduledDelay. >>>> scheduledDelay := nil ]. >>>> <snip> >>>> rest of method unchanged >>>> >>>> >>>> Then opened the Transcript and in Playground evaluated... >>>> Delay delaySchedulerClass: DelayMicrosecondScheduler. >>>> >>>> >>>> The Transcript results are shown below with some comments inserted. >>>> >>>> Now I guess the latency is affected by garbage collection. >>>> But one thing I was curious about is why the latency's were quantised in 1000s. >>>> >>>> Another interesting thing is that vast majority of the latency's were zero, >>>> which was a nice surprise, but can it be true? Or is it a consequence >>>> of the quantitisation rounding down? >>>> >>>> It seems that the idle-ness of the image affected how often a non-zero latency occurred. >>>> After I left the house for a while, the count of zero latency was very high, but a few still occurred. It would make sense there was less GC while idle. What is a good snippet of code to stress GC. I presume the latency might increase. >>>> >>>> >>>> zero latency count 2273 before latency 1000 >>>> zero latency count 943 before latency 1000 >>>> zero latency count 3666 before latency 1000 >>>> zero latency count 1643 before latency 1000 >>>> zero latency count 27 before latency 1000 >>>> "Left house for 20 minutes" >>>> zero latency count 12022 before latency 1000 >>>> zero latency count 15195 before latency 1000 >>>> zero latency count 41998 before latency 1000 >>>> "Returned from outing" >>>> zero latency count 128 before latency 1000 >>>> zero latency count 116 before latency 1000 >>>> zero latency count 555 before latency 1000 >>>> zero latency count 2377 before latency 1000 >>>> zero latency count 5423 before latency 1000 >>>> zero latency count 3178 before latency 1000 >>>> zero latency count 47 before latency 1000 >>>> zero latency count 2276 before latency 1000 >>>> "Left house to go shopping" >>>> zero latency count 6708 before latency 3000 >>>> zero latency count 4896 before latency 1000 >>>> zero latency count 433 before latency 1000 >>>> zero latency count 7106 before latency 1000 >>>> zero latency count 2195 before latency 1000 >>>> zero latency count 12397 before latency 1000 >>>> zero latency count 4815 before latency 2000 >>>> zero latency count 3480 before latency 1000 >>>> zero latency count 5690 before latency 1000 >>>> zero latency count 1670 before latency 1000 >>>> zero latency count 320 before latency 1000 >>>> zero latency count 1641 before latency 2000 >>>> zero latency count 290 before latency 1000 >>>> zero latency count 3835 before latency 1000 >>>> zero latency count 3818 before latency 1000 >>>> zero latency count 1132 before latency 1000 >>>> "Returned from shopping." >>>> zero latency count 2045 before latency 1000 >>>> zero latency count 671 before latency 1000 >>>> zero latency count 167 before latency 1000 >>>> zero latency count 147 before latency 1000 >>>> >>>> From the System Browser, inspecting the class variable LatencyCounts showed... >>>> 0 171329 >>>> 3000 1 >>>> 2000 3 >>>> 1000 49 >>> >>> P.S. after running a few more hours, LatencyCounts shows... >>> 0 1102691 >>> 1000 219 >>> 2000 12 >>> 3000 29 >>> 4000 21 >>> 5000 5 >>> 49000 1 >>> 75000 1 >> >> It's likely that the vm code is using the cheap microsecond click which is updated by the heartbeat, not the expensive one that interrogates the OS at the instant. If not, it's the OS. I'll check soon. > > I went looking and found an obvious source of quantitization multiplying by MicrosecondsPerMillisecond > in the Win32 currentUTCMicroseconds() line 109 here... > https://github.com/OpenSmalltalk/opensmalltalk-vm/blob/e17db79411cfec767e04… When that code was written there was no higher resolution alternative. But now I see that there is GetSystemTimePreciseAsFileTime as of Windows 8/Server 2012. Volunteers who would write in the alternative (by seeing if it exists in kernel32.dll and hence maintaining backward compatibility with Windows 7) are welcome. We could really do with this on Windows. > > Running the same experiment (for 20 mins) on Linux produced a latency distribution in line with what I was originally expecting... > usec count > 1 135 > 2 992 > 3 3250 > 4 21161 > 5 17309 > 6 5341 > 7 1230 > 8 313 > 9 79 > 10 44 > 11 46 > 12 50 > 13 36 > 14 58 > 15 73 > 16 38 > 17 12 > 18 7 > 19 5 > 20 1 > 21 3 > 23 2 > 24 4 > 25 4 > 26 5 > 27 4 > 29 3 > 30 6 > 31 5 > 32 6 > 33 5 > 34 2 > 35 2 > 36 5 > 37 4 > 39 4 > 40 2 > 41 2 > 44 3 > 46 2 > 47 1 > 48 4 > 49 3 > 51 2 > 53 1 > 54 1 > 56 2 > 58 1 > 59 1 > 61 1 > 66 2 > 70 1 > 82 1 > 86 1 > 184 1 > 544 1 > 576 1 > 1212 1 > 1402 1 > 2016 1 > > self average. "4.8"microseconds > self stdev. "12.9"microseconds > self average + (5 * self stdev) "69.1"microseconds > > So there are nine outliers outside five sigma. > @Clement, Would you expect your incremental GC to have much impact on those outliers? > or its target pause is ~1000usecs anyway? Not talking for ClÃ©ment, I think it would. But can your measurements correlate the outliers with GC pauses? Can you extend the experiment to interrogate the GC counts? For example, maintaining a WeakArray with: Object new, and observing when its element is nil is a cheap way of observing that a scavenge or GC has occurred, and a pin operation would ensure it was in old space. So you could allocate a new one to observe scavenges and pin another one, assigning a fresh Object new to it to observe full GCs. > cheers -ben

May 28, 2018

[Pharo 7.0-dev] Build #973: 22008-Move-GTPlaygroundBasicTest-cleaning-to-tearDown

by ci-pharo-ci-jenkins2＠inria.fr

There is a new Pharo build available! The status of the build #973 was: SUCCESS. The Pull Request #1451 was integrated: "22008-Move-GTPlaygroundBasicTest-cleaning-to-tearDown" Pull request url: https://github.com/pharo-project/pharo/pull/1451 Issue Url: https://pharo.fogbugz.com/f/cases/22008 Build Url: https://ci.inria.fr/pharo-ci-jenkins2/job/Test%20pending%20pull%20request%2…

May 28, 2018

[Pharo 7.0-dev] Build #972: 21815-shrink-for-OrderedCollection

by ci-pharo-ci-jenkins2＠inria.fr

There is a new Pharo build available! The status of the build #972 was: FAILURE. The Pull Request #1428 was integrated: "21815-shrink-for-OrderedCollection" Pull request url: https://github.com/pharo-project/pharo/pull/1428 Issue Url: https://pharo.fogbugz.com/f/cases/21815 Build Url: https://ci.inria.fr/pharo-ci-jenkins2/job/Test%20pending%20pull%20request%2…

May 28, 2018

Re: [Pharo-dev] IMPORTANT: ML Etiquette

by Esteban A. Maringolo

On 28/05/2018 12:36, Tim Mackinnon wrote: > Sent from my iPhone >> On 28 May 2018, at 17:26, Stephan Eggermont <stephan(a)stack.nl> wrote: >>> Sent from my iPhone >> That supports NewsTap, which makes it easy to not top-post. Iâm not sure >> why lines are not broken though. > > Iâll try my best - but the normal iPhone mail client simply doesnât work this way - it top posts. I'm fine with top posting. Even when a thread has both top and in-between replies. What really bothers me is that top posting "hides" the huge list of replies at the bottom. So a two line, top posted reply is a 500 lines of quoted email below. Top posting should be limited to one quoted email only. Otherwise you need to start referencing things by letters/numbers or better... reply below :) > I think weâre on a losing battle with this one. I agree. But I'm not giving it away without a fight :P Regards, -- Esteban A. Maringolo

May 28, 2018

[Pharo 7.0-dev] Build #971: Format methods from the suggestion menu in a debugger leads to a DNU â¦

by ci-pharo-ci-jenkins2＠inria.fr

There is a new Pharo build available! The status of the build #971 was: SUCCESS. The Pull Request #1423 was integrated: "Format methods from the suggestion menu in a debugger leads to a DNU â¦" Pull request url: https://github.com/pharo-project/pharo/pull/1423 Issue Url: https://pharo.fogbugz.com/f/cases/21377 Build Url: https://ci.inria.fr/pharo-ci-jenkins2/job/Test%20pending%20pull%20request%2…

May 28, 2018

Horrible mess in keyboard modifiers processing

by Pavel Krivanek

On Linux and Windows, the keyboard usage is very limited, especially in combination with the mouse. There is no way how to invoke any key corresponding to Mac command key (â). As result there are is no option how on these platforms select several nonadjacent methods in Calypso to display them at once and you cannot click on a method selector to reach the implementation. I tried to log modifiers on input events on all three platforms and this is the sad result: Mac: shift 0 1 1 0 ctrl 0 2 2 0 option (â¥) 0 4 4 0 command (â) 0 8 8 0 The numbers are in order: - modifier key pressed - mouse down - mouse up - modifier key released So on Mac the mouse events are processed correctly but the events related keyboard have no modifiers flags. Linux: shift 0 1 1 0 ctrl 0 0 2 0 win - 0 0 - alt 0 - - 0 altgr - 0 0 - So on Linux the Shift key is processed correctly, The Ctrl key does not have correct modifier flag when the mouse is pressed (which is strange because the combination of Shift+Ctrl returns 0 3 3 0). Win key changes click behavior to right mouse button. For Alt the mouse events are not detected at all and AltGr behaves as Win except the right mouse button simulation. The keyboard event modifiers are not set. Windows: shift 16 1 1 1 16 ctrl 17 2 2 2 17 win - 0 0 0 - alt 18 8 8 8 18 altgr 18 8 8 8 18 Windows as the only platform set modifiers for keyboard events. The mouse button release generates two events, Win key changes behavior to right mouse button, and Alt behaves the same way as AltGr. Cheers, -- Pavel

May 28, 2018

Re: [Pharo-dev] IMPORTANT: ML Etiquette

by Tim Mackinnon

Sent from my iPhone > On 28 May 2018, at 17:26, Stephan Eggermont <stephan(a)stack.nl> wrote: > > >> Sent from my iPhone > > That supports NewsTap, which makes it easy to not top-post. Iâm not sure > why lines are not broken though. Iâll try my best - but the normal iPhone mail client simply doesnât work this way - it top posts. I think weâre on a losing battle with this one. I do have a copy of NewsTap - Iâll try it again, but I recall mail is just more convenient. I think Stef is possibly right - if weâre arguing about top vs bottom - weâre better off writing code instead - it possibly means a thread has gone in too long. Tim

May 28, 2018

Re: [Pharo-dev] IMPORTANT: ML Etiquette

by Stephan Eggermont

Tim Mackinnon <tim(a)testit.works> wrote: > In a world of mobile phones, I think you have to accept top posting, itâs > just too hard to work otherwise on a small screen ... that said, quoting > a small piece to reply to - is doable, and I think we can all try to do that? I strongly prefer ML-etiquette. > Sent from my iPhone That supports NewsTap, which makes it easy to not top-post. Iâm not sure why lines are not broken though. Stephan

May 28, 2018