- Pharo-dev - lists.pharo.org

Implementing lazy loading for expandable collections (part 2)
by Max Leske Jan. 16, 2015

Jan. 16, 2015

Hi Alex I thought the following little experiment might interest you. I was trying to find out (as per your suggestion) how big the difference would be between: - setting the array in OrderedCollection to nil and initializing it to size 10 on first access - initializing the array in OrderedCollection to size 0. Expectation: #grow will take time, ergo, using an empty array should be slower than using a preallocated array. Experiment (Pharo 1.3 on SqueakVM): [ 10000 timesRepeat: [ | c | c := OrderedCollection new: 100000. 1 to: 10000 do: [ :i | c add: i ] ] ] timeToRun 17566 [ 10000 timesRepeat: [ | c | c := OrderedCollection new: 10000. 1 to: 10000 do: [ :i | c add: i ] ] ] timeToRun 19717 [ 10000 timesRepeat: [ | c | c := OrderedCollection new: 0. 1 to: 10000 do: [ :i | c add: i ] ] ] timeToRun 17598 [ 10000 timesRepeat: [ | c | c := OrderedCollection new: 1000. 1 to: 10000 do: [ :i | c add: i ] ] ] timeToRun 18084 Result: growing a collection is always faster (or at least nearly equally fast)! Explanation: What makes preinitialized collections so slow is the fact the the firstIndex variable will be set to the first third of the array (3333 in this case). When the last position in the array is reached, all elements are copied to the beginning of the array (2n operations). This copy operation is really slow (which can be seen from the first example with an oversized collection that does not require copying because the last position of the array is never reached). Experiment (Pharo 4 on PharoVM): (Iâm using bigger numbers here to compensate for the faster VM :) ) [ 10000 timesRepeat: [ | c | c := OrderedCollection new: 1000000. 1 to: 100000 do: [ :i | c add: i ] ] ] timeToRun "0:00:01:09.225" [ 10000 timesRepeat: [ | c | c := OrderedCollection new: 100000. 1 to: 100000 do: [ :i | c add: i ] ] ] timeToRun "0:00:00:37.653" [ 10000 timesRepeat: [ | c | c := OrderedCollection new: 0. 1 to: 100000 do: [ :i | c add: i ] ] ] timeToRun "0:00:00:40.152" [ 10000 timesRepeat: [ | c | c := OrderedCollection new: 10000. 1 to: 100000 do: [ :i | c add: i ] ] ] timeToRun "0:00:00:42.121â Result: growing a collection is easily fast enough, although minute performance gains are possible with preinitialized collections. Explanation: #makeRoomAtLast has been improved. Also: firstIndex is now set to the beginning of the array, preventing the copy operation that made the former measurements so slow. The run with the oversized collection is only so slow because it apparently takes a lot of time to assign the large array (?). Consequence of my little experiment: My initial implementation of the idea in you paper set the array in OrderedCollection to nil under the assumption that grow operations are very costly. That choice leads to many changes in methods of OrderedCollection where it is necessary to check if the array has been initialized yet. This introduces overhead and makes the code more error prone. I see now that it makes much more sense to use an empty array, the performance lost by growing ist neglectable. It is also interesting to note that there almost never seems to be a good reason to initialize an OrderedCollection with a large array. Cheers, Max

5 9

Object>>logCr ?
by Alexandre Bergel Jan. 16, 2015

Jan. 16, 2015

Hi! I like the new way to print things in the Transcript. But I am wondering, Object>>logCr self logCr: self printString why printString is used and not asString? âfooâ logCr prints âfooâ and not foo Cheers, Alexandre -- _,.;:~^~:;._,.;:~^~:;._,.;:~^~:;._,.;:~^~:;._,.;: Alexandre Bergel http://www.bergel.eu ^~:;._,.;:~^~:;._,.;:~^~:;._,.;:~^~:;._,.;:~^~:;.

3 2

Re: [Pharo-dev] Implementing lazy loading for expandable collections (part 2)
by Stephan Eggermont Jan. 16, 2015

Jan. 16, 2015

Eliot wrote: >About 90% of the entire time is in the string copying primitive. >Not much to be done there. Sista could speed this up by eliminating >the validation overhead of the primitive. But with our current VM technology >there's not a lot one can do given that OrderedCollection is organized as a flat stack. > This is probably a case where an ordered tree would be faster, albeit have a lot more space overhead. Switching collections backing store to a chunked representation once they grow beyond a certain threshold is probably the best thing to support larger collections. Stephan

1 0

Understanding Object Structures
by Eliot Miranda Jan. 16, 2015

Jan. 16, 2015

Hi All, On Jan 16, 2015, at 7:36 AM, Eliot Miranda <eliot.miranda(a)gmail.com> wrote: > Hi Marcel, > > On Jan 16, 2015, at 5:54 AM, Marcel Taeumel <marcel.taeumel(a)student.hpi.uni-potsdam.de> wrote: > >> Ah, okay. I was curious about the performance of iterating over an array vs. >> the linked list: >> >> a := Array new: 1000. >> l := LinkedList new. >> >> 1000 timesRepeat: [l add: (StackLink with: nil)]. >> >> [a do: [:ea | 1 + 2]] bench. "'94,600 per second.'" >> >> [l do: [:ea | 1 + 2]] bench. "'159,000 per second.'" >> >> I expected the opposite! :-O Why is that so? > > Look at the implementations of do:. > > In the Array case we iterate over the indices using SmallIntegers (fast arithmetic with no allocation) and dereference each element using the Object>>at: primitive, which (apart from type and bounds checking) is simple indexed memory access. So one read per element. > > In the LinkedList case each list element is fetched from the previous element and each value fetched from the element, so two reads per element. This is interesting. (Marcel, Chris, forgive me; I'm presuming; please don't take this personally). Marcel above appears to lack an intuition about the structure of Array vs LinkedList. And in developing a hash algorithm for a 32-bit subset of Floats a few weeks ago Chris appeared to lack an I tuition about Floats being boxed, assuming they were value types, not containers. As a VM implementer I carry around a clear picture (literally, I am a visual thinker) of objects in my head. Those pictures are key to my approach to design and optimization. I presume that for someone approaching the system given only a textual description of object structures, through class comments and method source it is difficult to develop a good picture or mental model. For me, I read the blue book first, which is replete with pictures. I know that historically visual inspector frameworks such as Jun have been able to auto-generate pictorial representations of specific object graphs. I wonder how useful it would be to provide support for designers to include pictorial representations in class comments. Again I presume that the text model would have to support inclusion of simple bitmaps (to avoid having to include a full drawing framework in the system) and that the designer would construct a sample graph, generate a diagram using a visual inspector framework using eg Jun, render it to a bitmap and include it in the class comment. A more elaborate system could of course include the sample graph and render it dynamically, that would allow exploration. Either approach would make an interesting project, yes? >> Best, >> Marcel >> >> >> >> -- >> View this message in context: http://forum.world.st/The-Trunk-Collections-mt-593-mcz-tp4799750p4799893.ht… >> Sent from the Squeak - Dev mailing list archive at Nabble.com. >>

3 3

[ANN] Pharo Consortium New Academic Partner: Ecole des Mines de Douai
by Marcus Denker Jan. 16, 2015

Jan. 16, 2015

The Pharo Consortium is very happy to announce that the Ecole des Mines de Douai has joined the Consortium as an Academic Partner. About - Ecole des Mines de Douai: http://www.mines-douai.fr - Ecole des Mines de Douai, IA: http://ia.mines-douai.fr - Pharo Consortium: http://consortium.pharo.org The goal of the Pharo Consortium is to allow companies and institutions to support the ongoing development and future of Pharo. Individuals can support Pharo via the Pharo Association: http://association.pharo.org

2 1

Any Deep Learning Algorithm related lib in Pharo ?
by CÃ©drick BÃ©ler Jan. 16, 2015

Jan. 16, 2015

Hi all, The question is in the title. Iâm just journeying in this field and was wandering if it exists something on deep learning algorithms. Some refs an introduction: http://www.toptal.com/machine-learning/an-introduction-to-deep-learning-fro… Python lib: http://www.pyimagesearch.com/2014/09/22/getting-started-deep-learning-pytho… http://deeplearning.net/tutorial/ Java lib (link in the tutorial) ... Cheers and best wishes all for 2015, CÃ©drick

2 1

Hosting the Pharo VM on MirageOS
by Joerg Beekmann, DeepCove Labs Jan. 16, 2015

Jan. 16, 2015

Hi I've been exploring the possibility of hosting the Squeak/Pharo VM on MirageOS. MirageOS is a Unikernal or "Library OS" where rather than creating an executable to be run on Linux the compiler analyses dependencies right down through the device drivers and creates a kernel that can be booted on Amazon AWS EC2. http://www.openmirage.org/ http://queue.acm.org/detail.cfm?id=2566628 http://blog.acolyer.org/2015/01/13/unikernels-library-operating-systems-for… It seems to me that having a Smalltalk hosed in this environment would be quite useful on a number of fronts: * A simple build and deploy of Smalltalk to the cloud. The Mirage group is building unikernals from sources and then because they are small committing the entire kernel to GitHub and then pushing the kernel to AWS EC2. Web sites and APIs are an obvious application. http://amirchaudhry.com/from-jekyll-to-unikernel-in-fifty-lines/. * Smalltalk on small devices. The Mirage group is running Xen with mirage on small Intel and ARM boards. See http://openmirage.org/blog/announcing-mirage-20-release/. * These kernels can be very small, on the order of 0.25Meg for minimal HTTP servers with boot times in the ms range. A project called Jitsu https://github.com/MagnusS/jitsu modifies a DNS server to boot kernels in response to socket requests. The system making the request is unaware of the boot. * The Mirage group envisions thousands of kernels running on a single Hypervisor with his speed inter-kernel communications. See http://openmirage.org/blog/update-on-vchan. The facilitates Smalltalk systems consisting of a swarm of communicating images where each image is single threaded and concurrency is via message passing. I've corresponded with the MirageOS group asking if they thought a language VM could be hosted on Mirage. The discussion is here: http://lists.xenproject.org/archives/html/mirageos-devel/2015-01/msg00053.h…. Assuming we understood each other it seems the answer is "yes that should be possible". My question is what would be involved on the Smalltalk side? I presume the effort is going to be mostly in the VM. Based on the conversation on the Mirage list I think the VM will need to be able to run a Library with an entry point called by Mirage. That seems similar to the " Embedding/VM as a DLL" project proposed here: http://www.mirandabanda.org/cogblog/cog-projects/. Is this project active? The comment indicates this is mostly refactoring & repackaging. Anyone have a perspective on this? Best regards Joerg

10 20

Finish one thing today
by Marcus Denker Jan. 16, 2015

Jan. 16, 2015

This is nice: http://finishonethingtoday.com Marcus

8 8

[pharo-project/pharo-core]
by GitHub Jan. 15, 2015

Jan. 15, 2015

Branch: refs/tags/40448 Home: https://github.com/pharo-project/pharo-core

1 0

[pharo-project/pharo-core] 814baf: 40448
by GitHub Jan. 15, 2015

Jan. 15, 2015

Branch: refs/heads/4.0 Home: https://github.com/pharo-project/pharo-core Commit: 814baf496e8f2042c42b6622b637c2aa214ad96b https://github.com/pharo-project/pharo-core/commit/814baf496e8f2042c42b6622… Author: Jenkins Build Server <board(a)pharo-project.org> Date: 2015-01-16 (Fri, 16 Jan 2015) Changed paths: M DebuggerModel.package/DebugSession.class/instance/debugging actions/stepInto_.st M DebuggerModel.package/DebugSession.class/instance/debugging actions/stepOver_.st M DebuggerModel.package/DebugSession.class/instance/debugging actions/stepThrough_.st M DebuggerModel.package/DebugSession.class/instance/evaluating/rewindContextToMethod_fromContext_.st M DebuggerModel.package/DebugSession.class/instance/evaluating/unwindAndRestartToContext_.st A DebuggerModel.package/DebugSession.class/instance/private/stepToFirstInterestingBytecodeIn_.st A Kernel.package/PackageManifest.class/README.md A Kernel.package/PackageManifest.class/class/testing/isManifest.st A Kernel.package/PackageManifest.class/definition.st A Kernel.package/PackageManifest.class/instance/code-critics/rejectClasses.st A Kernel.package/PackageManifest.class/instance/code-critics/rejectRules.st A Kernel.package/PackageManifest.class/instance/comparing/=.st A Kernel.package/PackageManifest.class/instance/comparing/hash.st R Manifest-Core.package/PackageManifest.class/README.md R Manifest-Core.package/PackageManifest.class/class/testing/isManifest.st R Manifest-Core.package/PackageManifest.class/definition.st R Manifest-Core.package/PackageManifest.class/instance/code-critics/rejectClasses.st R Manifest-Core.package/PackageManifest.class/instance/code-critics/rejectRules.st R Manifest-Core.package/PackageManifest.class/instance/comparing/=.st R Manifest-Core.package/PackageManifest.class/instance/comparing/hash.st A ScriptLoader40.package/ScriptLoader.class/instance/pharo - scripts/script448.st A ScriptLoader40.package/ScriptLoader.class/instance/pharo - updates/update40448.st M ScriptLoader40.package/ScriptLoader.class/instance/public/commentForCurrentUpdate.st A Traits.package/ClassTrait.class/instance/initialize-release/slots_.st A Traits.package/TApplyingOnClassSide.class/instance/initialize-release/slots_.st Log Message: ----------- 40448 14622 PackageManifest can not be in package Manifest-Core... https://pharo.fogbugz.com/f/cases/14622 14738 DebugSession assumes sending #stepToSendOrReturn is always needed https://pharo.fogbugz.com/f/cases/14738 14740 Faiing test: testMetaclassAndTraitClassRespectsPolymorphismRules https://pharo.fogbugz.com/f/cases/14740 http://files.pharo.org/image/40/40448.zip

1 0