performance 3.8 vs. 3.9a
Locked
 
|
Hi,
To me, 3.9a feels a little bit snappier than 3.8... (thanks to a lot of work by a lot of people, Eddie Cottongim I think did many of the big-gain changesets). So here is the benchmark again that opens 5*10 browsers on #abandonSources. This, of course, is an "apples vs. oranges" benchmark... e.g. there are many more systemcategories and classes in 3.9 than in 3.8... and early version don't render TrueType fonts for the window title. But in the end, we just want to open a Browser, and that should not take a too long. 3.5 5021.4 3.6 5282.2 3.7: 8838.4 3.8: 12010.8 3.9a: 7648.0 (Powerbook 1.5Ghz, all 32bit color, same vm) So at least, the trend of the "ever slower Squeak" has been stopped. 1.57 times faster then 3.8, faster then 3.7... but slower then 3.6. Nevertheless, not too bad, as a start. Marcus Code: (Dan's 10-Browser benchmark): ---------------------------------------------------- time _ 0. saveMorphs _ self currentWorld submorphs. 5 timesRepeat:[ self currentWorld removeAllMorphs. "heh, heh" time _ time + (Time millisecondsToRun: [ 1 to: 10 do: [:i | Browser fullOnClass: SystemDictionary selector: #abandonSources]. self currentWorld submorphs do: [:m | m delete. self currentWorld doOneCycle]]). ]. self currentWorld addAllMorphs: saveMorphs. time/5 asFloat "print it" |
|
|
Well let's see I noticed the default image 3.9a is set to 16bit,
change that to 32bit because 16bit requires an extra transformation step on the way to 32bit screens. 41% is interpreter loop 15% is in garbage collection and object allocation 5% is copyLooopNoSource 4% is copyLoopPixMap 4% is proceedRenderingScanline 4% is lookupMethodInClass Now if I alter things a bit and do Smalltalk setGCBiasToGrowGCLimit: 16*1024*1024. Smalltalk setGCBiasToGrow: 1. Then we shave 2% off the GC work and interpreter loop is now 43% 7006, with 3.8.11b4U vm on 1.5Ghz Pb 17 with other things running... 8536.0 16bit 8448.0 32bit 8349.8 32bit and GC tuning And about 30%+ is tied up in rendering then. I'll note in 3.8 a quick test shows times at 15451.6 with interpreter loop at 54% On 20-Mar-06, at 2:56 PM, Marcus Denker wrote: > Hi, > > To me, 3.9a feels a little bit snappier than 3.8... (thanks to a > lot of work by a lot > of people, Eddie Cottongim I think did many of the big-gain > changesets). > > So here is the benchmark again that opens 5*10 browsers on > #abandonSources. > > This, of course, is an "apples vs. oranges" benchmark... e.g. there > are many more > systemcategories and classes in 3.9 than in 3.8... and early > version don't render > TrueType fonts for the window title. But in the end, we just want > to open > a Browser, and that should not take a too long. > > 3.5 5021.4 > 3.6 5282.2 > 3.7: 8838.4 > 3.8: 12010.8 > 3.9a: 7648.0 > > (Powerbook 1.5Ghz, all 32bit color, same vm) > > So at least, the trend of the "ever slower Squeak" has been stopped. > 1.57 times faster then 3.8, faster then 3.7... but slower then 3.6. > Nevertheless, not too bad, as a start. > > Marcus ======================================================================== === John M. McIntosh <[hidden email]> 1-800-477-2659 Corporate Smalltalk Consulting Ltd. http://www.smalltalkconsulting.com ======================================================================== === |
|
|
On 21.03.2006, at 01:48, John M McIntosh wrote: > Well let's see I noticed the default image 3.9a is set to 16bit, > change that to 32bit because 16bit requires an extra transformation > step on > the way to 32bit screens. Yes, I did the test already with 32 bits (this will be default starting 7015... it's 2006 after all...). > > 41% is interpreter loop > 15% is in garbage collection and object allocation > 5% is copyLooopNoSource > 4% is copyLoopPixMap > 4% is proceedRenderingScanline > 4% is lookupMethodInClass > > Now if I alter things a bit and do > > Smalltalk setGCBiasToGrowGCLimit: 16*1024*1024. > Smalltalk setGCBiasToGrow: 1. > Would it make sense to set to be the default in 3.9a? marcus |
|
|
On 20-Mar-06, at 5:24 PM, Marcus Denker wrote: >> > > Would it make sense to set to be the default in 3.9a? Likely, but awaiting support from other VM... > > marcus -- ======================================================================== === John M. McIntosh <[hidden email]> 1-800-477-2659 Corporate Smalltalk Consulting Ltd. http://www.smalltalkconsulting.com ======================================================================== === |
Re: performance 3.8 vs. 3.9a
|
|
What do you imply? that it will never happen?
On 21 mars 06, at 03:41, John M McIntosh wrote: > > On 20-Mar-06, at 5:24 PM, Marcus Denker wrote: > >>> >> >> Would it make sense to set to be the default in 3.9a? > > Likely, but awaiting support from other VM... > >> >> marcus > > -- > ====================================================================== > ===== > John M. McIntosh <[hidden email]> 1-800-477-2659 > Corporate Smalltalk Consulting Ltd. http:// > www.smalltalkconsulting.com > ====================================================================== > ===== > > |
Re: performance 3.8 vs. 3.9a
|
|
In reply to this post by Marcus Denker
cool
I think that we should learn from this category speedup. 1- It seems to me that lot of the design choices which made absolute sense in the early life of Smalltalk (the way system category,class category for example are managed) because running something on that machines was a real challenge, now should be revised. 2- lot of small changes can make a difference. Incremental process is good even if there is no problem throwing away them when a better solution arrives. So I encourage everybody to help :) For example the network seems a bit requiring help Stef On 20 mars 06, at 23:56, Marcus Denker wrote: > Hi, > > To me, 3.9a feels a little bit snappier than 3.8... (thanks to a > lot of work by a lot > of people, Eddie Cottongim I think did many of the big-gain > changesets). > > So here is the benchmark again that opens 5*10 browsers on > #abandonSources. > > This, of course, is an "apples vs. oranges" benchmark... e.g. there > are many more > systemcategories and classes in 3.9 than in 3.8... and early > version don't render > TrueType fonts for the window title. But in the end, we just want > to open > a Browser, and that should not take a too long. > > 3.5 5021.4 > 3.6 5282.2 > 3.7: 8838.4 > 3.8: 12010.8 > 3.9a: 7648.0 > > (Powerbook 1.5Ghz, all 32bit color, same vm) > > So at least, the trend of the "ever slower Squeak" has been stopped. > 1.57 times faster then 3.8, faster then 3.7... but slower then 3.6. > Nevertheless, not too bad, as a start. > > Marcus > > Code: (Dan's 10-Browser benchmark): > ---------------------------------------------------- > > time _ 0. > saveMorphs _ self currentWorld submorphs. > 5 timesRepeat:[ > self currentWorld removeAllMorphs. "heh, heh" > time _ time + (Time millisecondsToRun: > [ > 1 to: 10 do: [:i | Browser fullOnClass: SystemDictionary selector: > #abandonSources]. > self currentWorld submorphs do: [:m | m delete. self currentWorld > doOneCycle]]). > ]. > self currentWorld addAllMorphs: saveMorphs. > time/5 asFloat "print it" |
|
|
In reply to this post by stéphane ducasse-2
I suspect they run into edge cases with the old solution in Croquet,
so turning this feature on is likely, however someone has to build a new unix and windows vm first... On 20-Mar-06, at 11:37 PM, stéphane ducasse wrote: > What do you imply? that it will never happen? > > On 21 mars 06, at 03:41, John M McIntosh wrote: > >> >> On 20-Mar-06, at 5:24 PM, Marcus Denker wrote: >> >>>> >>> >>> Would it make sense to set to be the default in 3.9a? >> >> Likely, but awaiting support from other VM... -- ======================================================================== === John M. McIntosh <[hidden email]> 1-800-477-2659 Corporate Smalltalk Consulting Ltd. http://www.smalltalkconsulting.com ======================================================================== === |
«
Return to Squeak - Dev
|
1 view|%1 views
| Free forum by Nabble | Edit this page |