[Pharo-project] Pharo + Exupery VM + Something else than Accuny
Hi, I'm playing with the Exupery VM on Pharo. If I change the font to something else than Accuny, the underlined texts are not underlined anymore. To reproduce: - download and use http://ftp.squeak.org/Exupery/vms/exupery-vm-0.15-linux.tz - go to preferences->fonts->list fonts - select Bistream Charter/regular/10, and apply - open monticello and open the Pharo repository You will see that the installed packages are not underlined anymore. Switch back to Accuny to see them underlined. -- Damien Cassou http://damiencassou.seasidehosting.st
On Sat, Mar 7, 2009 at 11:57 AM, Hilaire Fernandes <hilaire@ofset.org> wrote:
And what about the speed improvment ?
Should there be any? I haven't played enough with it. If you are talking about the closures, I don't think I will see any improvement until the blocks are converted and I haven't done that. -- Damien Cassou http://damiencassou.seasidehosting.st
I am talking about performance, benchmark, etc. Is it not Exupery about improving the speed of the VM? 2009/3/7 Damien Cassou <damien.cassou@gmail.com>:
On Sat, Mar 7, 2009 at 11:57 AM, Hilaire Fernandes <hilaire@ofset.org> wrote:
And what about the speed improvment ?
Should there be any? I haven't played enough with it. If you are talking about the closures, I don't think I will see any improvement until the blocks are converted and I haven't done that.
-- Damien Cassou http://damiencassou.seasidehosting.st
_______________________________________________ Pharo-project mailing list Pharo-project@lists.gforge.inria.fr http://lists.gforge.inria.fr/cgi-bin/mailman/listinfo/pharo-project
Hilaire Fernandes writes:
I am talking about performance, benchmark, etc. Is it not Exupery about improving the speed of the VM?
Exupery is, these VMs include the support Exupery needs but the compiler is in the image. If you load and start Exupery you may see some gains. I haven't yet tuned the background compiler for practical use, it currently will compile every method it sees which is great for debugging but not good for providing a practical speed improvement. Compiled code takes up more cache than bytecode so there can be a loss due to compiling methods that are hardly ever used. Here's the current benchmarks: arithmaticLoopBenchmark 390 compiled 80 ratio: 4.875 bytecodeBenchmark 724 compiled 250 ratio: 2.895 sendBenchmark 663 compiled 385 ratio: 1.722 doLoopsBenchmark 381 compiled 235 ratio: 1.621 pointCreation 394 compiled 389 ratio: 1.013 largeExplorers 269 compiled 210 ratio: 1.280 compilerBenchmark 273 compiled 250 ratio: 1.092 Cumulative Time 413.408 compiled 232.706 ratio 1.777 With these VMs you could load Exupery from either SqueakMap or Universes and start it. Documentation is here: http://wiki.squeak.org/squeak/Exupery I may add full method inlining next rather than tune for practical performance. Full method inlining will speed up common sends extensively. In summary, there may be a practical performance gain now or there may not but Exupery is getting close to being seriously useful. Bryce
Bryce, I'll play dumb here - actually, I'm not playing in this case :) The little I know about Exupery comes from FAQs. In the past at least, you cautioned against enabling dynamic compiling of methods, at least in production. How does that square with full-inlining? Are they different concepts, or have thing improved enough that you are now confident enough to throw the switch? Full-inlining sounds like it could consume a lot of space. Is that a concern, or does the selection of what it compiles keep it manageable? How does it decide what to compile, and (more relevant to my dumb questions) how does that differ from the dynamic compilation? Please don't feel attacked - I'm simply trying to understand what is hopefully just around the corner. Bill -----Original Message----- From: pharo-project-bounces@lists.gforge.inria.fr [mailto:pharo-project-bounces@lists.gforge.inria.fr] On Behalf Of bryce@kampjes.demon.co.uk Sent: Saturday, March 07, 2009 6:55 AM To: Pharo-project@lists.gforge.inria.fr Subject: {Spam?} Re: [Pharo-project] Pharo + Exupery VM + Something else than Accuny Hilaire Fernandes writes:
I am talking about performance, benchmark, etc. Is it not Exupery about improving the speed of the VM?
Exupery is, these VMs include the support Exupery needs but the compiler is in the image. If you load and start Exupery you may see some gains. I haven't yet tuned the background compiler for practical use, it currently will compile every method it sees which is great for debugging but not good for providing a practical speed improvement. Compiled code takes up more cache than bytecode so there can be a loss due to compiling methods that are hardly ever used. Here's the current benchmarks: arithmaticLoopBenchmark 390 compiled 80 ratio: 4.875 bytecodeBenchmark 724 compiled 250 ratio: 2.895 sendBenchmark 663 compiled 385 ratio: 1.722 doLoopsBenchmark 381 compiled 235 ratio: 1.621 pointCreation 394 compiled 389 ratio: 1.013 largeExplorers 269 compiled 210 ratio: 1.280 compilerBenchmark 273 compiled 250 ratio: 1.092 Cumulative Time 413.408 compiled 232.706 ratio 1.777 With these VMs you could load Exupery from either SqueakMap or Universes and start it. Documentation is here: http://wiki.squeak.org/squeak/Exupery I may add full method inlining next rather than tune for practical performance. Full method inlining will speed up common sends extensively. In summary, there may be a practical performance gain now or there may not but Exupery is getting close to being seriously useful. Bryce _______________________________________________ Pharo-project mailing list Pharo-project@lists.gforge.inria.fr http://lists.gforge.inria.fr/cgi-bin/mailman/listinfo/pharo-project
Schwab,Wilhelm K writes:
Bryce,
I'll play dumb here - actually, I'm not playing in this case :)
The little I know about Exupery comes from FAQs. In the past at least, you cautioned against enabling dynamic compiling of methods, at least in production. How does that square with full-inlining? Are they different concepts, or have thing improved enough that you are now confident enough to throw the switch? Full-inlining sounds like it could consume a lot of space. Is that a concern, or does the selection of what it compiles keep it manageable? How does it decide what to compile, and (more relevant to my dumb questions) how does that differ from the dynamic compilation?
Please don't feel attacked - I'm simply trying to understand what is hopefully just around the corner.
What's changed is I've fixed a few bugs. It's possible that Exupery is reliable. I've yet done enough testing to be sure but since the last fix I haven't yet seen Exupery cause a crash. I'm confident enough to throw the switch. By full inlining all I mean is the ability to inline entire methods. Exupery will already dynamically inline the primitives it compiles to machine code. Dynamic inlining is the critical optimisation to allow further optimisations, without it the compiler can only see a method or a block at a time. It can't see where the loops are as they're probably in a different method as with #do: loops. This really limits the optimisations that can be performed later. Full inlining will be driven by profiling so it's only the hot spots that should get inlined, at least after I've tuned the profiler. Inlining can sometimes reduce the amount of code executed. For small methods it's easy to spend more code on the message send than is consumed by the method body. Initially it's likely the profiler will make bad inlining decisions like the current profiler will compile every method seen eventually which is probably bad for performance, compiling everything at least doesn't avoid compiling infrequent methods that may trigger a compiler bug. Full method inlining is a form of dynamic compilation. It lets the dynamic compiler inline entire methods when it sees the need. Bryce.
2009/3/7 <bryce@kampjes.demon.co.uk>:
With these VMs you could load Exupery from either SqueakMap or Universes and start it. Documentation is here:
I installed exupery code in pharo 203, run the test, got 4 errors on #testBlockBug3, #testBlockNonLocalReturnRecycledContext, #testDelayWaitStreeTest, #testStressFailure1 Next I did three benchmarks on 3 different setups with DrGeo MessageTally spyOn [Carre new]. Carre new instanciate a DrGeo canvas with Smalltalk programmed figure, it results in a self-repeating sketch (see screenshot) 1. image with exupery runned with exupery VM 2. image without exupery runned with stock 3.9 VM 3. image without exupery runned with exupery VM I used exupery image Damien pointed to in his email. Do you want the images ? Hilaire
And the fourth benchmark with compiled methods 2009/3/7 Hilaire Fernandes <hilaire@ofset.org>:
2009/3/7 Â <bryce@kampjes.demon.co.uk>:
With these VMs you could load Exupery from either SqueakMap or Universes and start it. Documentation is here:
I installed exupery code in pharo 203, run the test, got 4 errors on #testBlockBug3, #testBlockNonLocalReturnRecycledContext, #testDelayWaitStreeTest, #testStressFailure1
Next I did three benchmarks on 3 different setups with DrGeo MessageTally spyOn [Carre new].
Carre new instanciate a DrGeo canvas with Smalltalk programmed figure, it results in a self-repeating sketch (see screenshot)
1. image with exupery runned with exupery VM 2. image without exupery runned with stock 3.9 VM 3. image without exupery runned with exupery VM
I used exupery image Damien pointed to in his email.
Do you want the images ?
Hilaire
Next I tried the Exupery profiler, but still the benchmark is not as good as without the compiler. Hilaire 2009/3/7 Hilaire Fernandes <hilaire@ofset.org>:
And the fourth benchmark with compiled methods
2009/3/7 Hilaire Fernandes <hilaire@ofset.org>:
2009/3/7 Â <bryce@kampjes.demon.co.uk>:
With these VMs you could load Exupery from either SqueakMap or Universes and start it. Documentation is here:
I installed exupery code in pharo 203, run the test, got 4 errors on #testBlockBug3, #testBlockNonLocalReturnRecycledContext, #testDelayWaitStreeTest, #testStressFailure1
Next I did three benchmarks on 3 different setups with DrGeo MessageTally spyOn [Carre new].
Carre new instanciate a DrGeo canvas with Smalltalk programmed figure, it results in a self-repeating sketch (see screenshot)
1. image with exupery runned with exupery VM 2. image without exupery runned with stock 3.9 VM 3. image without exupery runned with exupery VM
I used exupery image Damien pointed to in his email.
Do you want the images ?
Hilaire
Hilaire Fernandes writes:
2009/3/7 <bryce@kampjes.demon.co.uk>:
With these VMs you could load Exupery from either SqueakMap or Universes and start it. Documentation is here:
I installed exupery code in pharo 203, run the test, got 4 errors on #testBlockBug3, #testBlockNonLocalReturnRecycledContext, #testDelayWaitStreeTest, #testStressFailure1
Next I did three benchmarks on 3 different setups with DrGeo MessageTally spyOn [Carre new].
Carre new instanciate a DrGeo canvas with Smalltalk programmed figure, it results in a self-repeating sketch (see screenshot)
1. image with exupery runned with exupery VM 2. image without exupery runned with stock 3.9 VM 3. image without exupery runned with exupery VM
I used exupery image Damien pointed to in his email.
Do you want the images ?
Thanks, I'd like to look at the images later but probably not until after the next release. The decision to work on full method inlining is a decision to not do much tuning until after a critical optimisation is in place. It is possible that running Exupery can slow down code. Normally this is because it either didn't compile something it needed to for a given benchmark. In your case it's suspicious that it's only compiled one of the leaf methods. It's also possible that I need to optimise something where Exupery is slower than the interpreter, I'm not aware of any such cases but further profiling may find some. Bryce
It will be great. Hilaire 2009/3/7 <bryce@kampjes.demon.co.uk>:
Hilaire Fernandes writes:  > 2009/3/7  <bryce@kampjes.demon.co.uk>:  >  > > With these VMs you could load Exupery from either SqueakMap  > > or Universes and start it. Documentation is here:  >  > I installed exupery code in pharo 203, run the test, got 4 errors on  > #testBlockBug3, #testBlockNonLocalReturnRecycledContext,  > #testDelayWaitStreeTest, #testStressFailure1  >  > Next I did three benchmarks on 3 different setups with DrGeo  > MessageTally spyOn [Carre new].  >  > Carre new instanciate a DrGeo canvas with Smalltalk programmed figure,  > it results in a self-repeating sketch (see screenshot)  >  > 1. image with exupery runned with exupery VM  > 2. image without exupery runned with stock 3.9 VM  > 3. image without exupery runned with exupery VM  >  > I used exupery image Damien pointed to in his email.  >  > Do you want the images ?
Thanks, I'd like to look at the images later but probably not until after the next release. The decision to work on full method inlining is a decision to not do much tuning until after a critical optimisation is in place.
It is possible that running Exupery can slow down code. Normally this is because it either didn't compile something it needed to for a given benchmark. In your case it's suspicious that it's only compiled one of the leaf methods. It's also possible that I need to optimise something where Exupery is slower than the interpreter, I'm not aware of any such cases but further profiling may find some.
Bryce
_______________________________________________ Pharo-project mailing list Pharo-project@lists.gforge.inria.fr http://lists.gforge.inria.fr/cgi-bin/mailman/listinfo/pharo-project
participants (4)
-
bryce＠kampjes.demon.co.uk -
Damien Cassou -
Hilaire Fernandes -
Schwab,Wilhelm K