Hi,
Since the work on the integration of PPContext in PetitParser, there is significant performance degradation. I have already mentioned that on simple grammar the factor is about 2. But on complex grammar (for example, our proprietary parser for 4D language ), we have seen that degradation is goes well beyond this factor. So, for example, for 800K lines that we parse in under 10 minutes without PPContext work, with the latest version it goes beyond 2h. I have not been able to reproduce my case on simple grammars. May be we can have some benchmarks using open source parsers and large code bases (e.g. PP Java Parser on JHotDraw or ArgoUML). Currently, I circumvent this issue by using PP v1.51 but I can provide relevant feedback and run benches on any improvements. regards. Usman _______________________________________________ Moose-dev mailing list [hidden email] https://www.iam.unibe.ch/mailman/listinfo/moose-dev |
Hi, Hmm. I thought this was fixed and that you said at the end that the performance penalty is no longer a factor of 2. Do you have a simple example for checking the simple case of 2x performance loss? Doru On Fri, Jan 30, 2015 at 12:28 PM, Usman Bhatti <[hidden email]> wrote:
_______________________________________________ Moose-dev mailing list [hidden email] https://www.iam.unibe.ch/mailman/listinfo/moose-dev |
Here is a script that shows the factor of 2 in time taken for parsing. The idea of the script is to introduce several wrong branches before hitting the correct one (originally proposed by Guillaume).
|rule wrongBranches| [wrongBranches := (25 to: 45) inject: 24 asCharacter asParser into: [:acc :each | acc / each asCharacter asParser]. rule := PPDelegateParser new. rule setParser: $a asParser / ((wrongBranches / $. asParser), rule). rule parse: (String streamContents: [ :s | 30000 timesRepeat: [s nextPut: $.]. s nextPut: $a ])] timeToRun Below are the graphs that were generated when varying the string length provided to the above parser description. Green part shows the time taken by the older version and the red part with the latest PP. Data values and graph generated by the script in the attached file. regards. usman On Fri, Jan 30, 2015 at 1:30 PM, Tudor Girba <[hidden email]> wrote:
_______________________________________________ Moose-dev mailing list [hidden email] https://www.iam.unibe.ch/mailman/listinfo/moose-dev script-graph.txt (1K) Download Attachment |
Usman, this is a really cool bug report !
> On 30 Jan 2015, at 16:36, Usman Bhatti <[hidden email]> wrote: > > Here is a script that shows the factor of 2 in time taken for parsing. The idea of the script is to introduce several wrong branches before hitting the correct one (originally proposed by Guillaume). > > |rule wrongBranches| > [wrongBranches := (25 to: 45) inject: 24 asCharacter asParser into: [:acc :each | acc / each asCharacter asParser]. > rule := PPDelegateParser new. > rule setParser: $a asParser / ((wrongBranches / $. asParser), rule). > rule parse: (String streamContents: [ :s | > 30000 timesRepeat: [s nextPut: $.]. s nextPut: $a ])] timeToRun > > > Below are the graphs that were generated when varying the string length provided to the above parser description. Green part shows the time taken by the older version and the red part with the latest PP. > > <graph.png> > > Data values and graph generated by the script in the attached file. > > regards. > > usman > > On Fri, Jan 30, 2015 at 1:30 PM, Tudor Girba <[hidden email]> wrote: > Hi, > > Hmm. I thought this was fixed and that you said at the end that the performance penalty is no longer a factor of 2. > > Do you have a simple example for checking the simple case of 2x performance loss? > > Doru > > > On Fri, Jan 30, 2015 at 12:28 PM, Usman Bhatti <[hidden email]> wrote: > Hi, > > Since the work on the integration of PPContext in PetitParser, there is significant performance degradation. I have already mentioned that on simple grammar the factor is about 2. But on complex grammar (for example, our proprietary parser for 4D language ), we have seen that degradation is goes well beyond this factor. So, for example, for 800K lines that we parse in under 10 minutes without PPContext work, with the latest version it goes beyond 2h. > > I have not been able to reproduce my case on simple grammars. May be we can have some benchmarks using open source parsers and large code bases (e.g. PP Java Parser on JHotDraw or ArgoUML). > > Currently, I circumvent this issue by using PP v1.51 but I can provide relevant feedback and run benches on any improvements. > > regards. > > Usman > > _______________________________________________ > Moose-dev mailing list > [hidden email] > https://www.iam.unibe.ch/mailman/listinfo/moose-dev > > > > > -- > www.tudorgirba.com > > "Every thing has its own flow" > > _______________________________________________ > Moose-dev mailing list > [hidden email] > https://www.iam.unibe.ch/mailman/listinfo/moose-dev > > > <script-graph.txt>_______________________________________________ > Moose-dev mailing list > [hidden email] > https://www.iam.unibe.ch/mailman/listinfo/moose-dev _______________________________________________ Moose-dev mailing list [hidden email] https://www.iam.unibe.ch/mailman/listinfo/moose-dev |
On Fri, Jan 30, 2015 at 9:36 PM, Sven Van Caekenberghe <[hidden email]> wrote: Usman, this is a really cool bug report ! Thanks :)
_______________________________________________ Moose-dev mailing list [hidden email] https://www.iam.unibe.ch/mailman/listinfo/moose-dev |
Hey all, thanks for the report, I will have a look at it asap. Cheers Jan On 1 Feb 2015 20:55, "Usman Bhatti" <[hidden email]> wrote:
_______________________________________________ Moose-dev mailing list [hidden email] https://www.iam.unibe.ch/mailman/listinfo/moose-dev |
Hey all, Please note that a new version in of PetitParser (JanKurs-271) should improve performance roughly comparable to the original performance. On my computer, I got improvement from 3900ms to 1300ms. On 1 February 2015 at 22:55, Jan Kurš <[hidden email]> wrote:
_______________________________________________ Moose-dev mailing list [hidden email] https://www.iam.unibe.ch/mailman/listinfo/moose-dev |
Great news. @Usman: could you check? Doru On Fri, Feb 20, 2015 at 9:04 AM, Jan Kurš <[hidden email]> wrote:
_______________________________________________ Moose-dev mailing list [hidden email] https://www.iam.unibe.ch/mailman/listinfo/moose-dev |
I executed the script that I shared earlier in this thread and I see a considerable performance improvement in the latest version of PP. The attached chart shows the comparison chart between PP v 1.51 (green) with the newer version (red). For shorter strings (i.e. to the left on x-axis), the execution time is almost comparable to the previous version of PetitParser that dates before changes introduced by Jan.
However, when I take the new version and parse the code source of larger code bases, the execution time still get a hit. For example, parsing 100K lines takes almost twice longer. The increase is exponential with even larger ones. That is consistent with results we see in the graph (the larger the chain to be parsed i.e. going from left to right in the x-axis, the higher the performance penalty). Meaning I still prefer the older version. What will be good is to prepare some benches with Java Parser in PP, as it is open source, and a set of programs (e.g. ArguUML) that can serve as a standard to test execution time for PP that we can run as a jenkins service. regards. On Fri, Feb 20, 2015 at 9:09 AM, Tudor Girba <[hidden email]> wrote:
_______________________________________________ Moose-dev mailing list [hidden email] https://www.iam.unibe.ch/mailman/listinfo/moose-dev |
Hey, I was not aware about some exponential element in the overall complexity, it is strange and I will investigate it. Thank you Usman, for pointing this out. In general, it is hard to add features into the PetitParser and preserve performance. Therefore, we work on a tool, that will take PetitParser, analyze it and generates faster parser. The idea is to use PetitParser while developing the grammar and than generate a fast parser for "real" use. We hope we will be present the tool during the next ESUG. Cheers, Jan On 20 February 2015 at 17:47, Usman Bhatti <[hidden email]> wrote:
_______________________________________________ Moose-dev mailing list [hidden email] https://www.iam.unibe.ch/mailman/listinfo/moose-dev |
Le 21/02/2015 11:14, Jan Kurš a écrit :
> Hey, > > I was not aware about some exponential element in the overall > complexity, it is strange and I will investigate it. Thank you Usman, > for pointing this out. > > In general, it is hard to add features into the PetitParser and preserve > performance. Therefore, we work on a tool, that will take PetitParser, > analyze it and generates faster parser. The idea is to use PetitParser > while developing the grammar and than generate a fast parser for "real" > use. We hope we will be present the tool during the next ESUG. I certainly be interested with what you come up with. From profiling PetitParser, it seems there is at least one low-hanging fruit in term of performance. And thanks too Usman, for pointing out long source code files may (or do) show diverging asymptotic behaviors. And real life use for most programming languages are very long files :( Thierry > > Cheers, > Jan > > On 20 February 2015 at 17:47, Usman Bhatti <[hidden email] > <mailto:[hidden email]>> wrote: > > I executed the script that I shared earlier in this thread and I see > a considerable performance improvement in the latest version of > PP. The attached chart shows the comparison chart between PP v 1.51 > (green) with the newer version (red). For shorter strings (i.e. to > the left on x-axis), the execution time is almost comparable to the > previous version of PetitParser that dates before changes introduced > by Jan. > > However, when I take the new version and parse the code source of > larger code bases, the execution time still get a hit. For example, > parsing 100K lines takes almost twice longer. The increase is > exponential with even larger ones. That is consistent with results > we see in the graph (the larger the chain to be parsed i.e. going > from left to right in the x-axis, the higher the performance > penalty). Meaning I still prefer the older version. > > What will be good is to prepare some benches with Java Parser in PP, > as it is open source, and a set of programs (e.g. ArguUML) that can > serve as a standard to test execution time for PP that we can run as > a jenkins service. > > > Inline image 1 > > regards. > > On Fri, Feb 20, 2015 at 9:09 AM, Tudor Girba <[hidden email] > <mailto:[hidden email]>> wrote: > > Great news. > > @Usman: could you check? > > Doru > > On Fri, Feb 20, 2015 at 9:04 AM, Jan Kurš <[hidden email] > <mailto:[hidden email]>> wrote: > > Hey all, Please note that a new version in of PetitParser > (JanKurs-271) should improve performance roughly comparable > to the original performance. On my computer, I got > improvement from 3900ms to 1300ms. > > > Cheers, > Jan > > On 1 February 2015 at 22:55, Jan Kurš <[hidden email] > <mailto:[hidden email]>> wrote: > > Hey all, thanks for the report, I will have a look at it > asap. > > Cheers Jan > > On 1 Feb 2015 20:55, "Usman Bhatti" > <[hidden email] <mailto:[hidden email]>> > wrote: > > > > On Fri, Jan 30, 2015 at 9:36 PM, Sven Van > Caekenberghe <[hidden email] <mailto:[hidden email]>> wrote: > > Usman, this is a really cool bug report ! > > > Thanks :) > > > > On 30 Jan 2015, at 16:36, Usman Bhatti <[hidden email] > <mailto:[hidden email]>> wrote: > > > > Here is a script that shows the factor of 2 in time taken for parsing. The idea of the script is to introduce several wrong branches before hitting the correct one (originally proposed by Guillaume). > > > > |rule wrongBranches| > > [wrongBranches := (25 to: 45) inject: 24 asCharacter asParser into: [:acc :each | acc / each asCharacter asParser]. > > rule := PPDelegateParser new. > > rule setParser: $a asParser / ((wrongBranches / $. asParser), rule). > > rule parse: (String streamContents: [ :s | > > 30000 timesRepeat: [s nextPut: $.]. s nextPut: $a ])] timeToRun > > > > > > Below are the graphs that were generated when varying the string length provided to the above parser description. Green part shows the time taken by the older version and the red part with the latest PP. > > > > <graph.png> > > > > Data values and graph generated by the script > in the attached file. > > > > regards. > > > > usman > > > > On Fri, Jan 30, 2015 at 1:30 PM, Tudor Girba > <[hidden email] > <mailto:[hidden email]>> wrote: > > Hi, > > > > Hmm. I thought this was fixed and that you > said at the end that the performance penalty is > no longer a factor of 2. > > > > Do you have a simple example for checking the > simple case of 2x performance loss? > > > > Doru > > > > > > On Fri, Jan 30, 2015 at 12:28 PM, Usman > Bhatti <[hidden email] > <mailto:[hidden email]>> wrote: > > Hi, > > > > Since the work on the integration of > PPContext in PetitParser, there is significant > performance degradation. I have already > mentioned that on simple grammar the factor is > about 2. But on complex grammar (for example, > our proprietary parser for 4D language ), we > have seen that degradation is goes well beyond > this factor. So, for example, for 800K lines > that we parse in under 10 minutes without > PPContext work, with the latest version it goes > beyond 2h. > > > > I have not been able to reproduce my case on > simple grammars. May be we can have some > benchmarks using open source parsers and large > code bases (e.g. PP Java Parser on JHotDraw or > ArgoUML). > > > > Currently, I circumvent this issue by using > PP v1.51 but I can provide relevant feedback and > run benches on any improvements. > > > > regards. > > > > Usman > > > > _______________________________________________ > > Moose-dev mailing list > > [hidden email] > <mailto:[hidden email]> > > > https://www.iam.unibe.ch/mailman/listinfo/moose-dev > > > > > > > > > > -- > > www.tudorgirba.com <http://www.tudorgirba.com> > > > > "Every thing has its own flow" > > > > _______________________________________________ > > Moose-dev mailing list > > [hidden email] > <mailto:[hidden email]> > > > https://www.iam.unibe.ch/mailman/listinfo/moose-dev > > > > > > > <script-graph.txt>_______________________________________________ > > Moose-dev mailing list > > [hidden email] > <mailto:[hidden email]> > > > https://www.iam.unibe.ch/mailman/listinfo/moose-dev > > > _______________________________________________ > Moose-dev mailing list > [hidden email] > <mailto:[hidden email]> > https://www.iam.unibe.ch/mailman/listinfo/moose-dev > > > > _______________________________________________ > Moose-dev mailing list > [hidden email] <mailto:[hidden email]> > https://www.iam.unibe.ch/mailman/listinfo/moose-dev > > > > _______________________________________________ > Moose-dev mailing list > [hidden email] <mailto:[hidden email]> > https://www.iam.unibe.ch/mailman/listinfo/moose-dev > > > > > -- > www.tudorgirba.com <http://www.tudorgirba.com> > > "Every thing has its own flow" > > _______________________________________________ > Moose-dev mailing list > [hidden email] <mailto:[hidden email]> > https://www.iam.unibe.ch/mailman/listinfo/moose-dev > > > > _______________________________________________ > Moose-dev mailing list > [hidden email] <mailto:[hidden email]> > https://www.iam.unibe.ch/mailman/listinfo/moose-dev > > > > > _______________________________________________ > Moose-dev mailing list > [hidden email] > https://www.iam.unibe.ch/mailman/listinfo/moose-dev > _______________________________________________ Moose-dev mailing list [hidden email] https://www.iam.unibe.ch/mailman/listinfo/moose-dev |
Hi,
On Sat, Feb 21, 2015 at 8:07 PM, Thierry Goubier <[hidden email]> wrote: Le 21/02/2015 11:14, Jan Kurš a écrit : What is that log hanging fruit? Doru And thanks too Usman, for pointing out long source code files may (or do) show diverging asymptotic behaviors. And real life use for most programming languages are very long files :( _______________________________________________ Moose-dev mailing list [hidden email] https://www.iam.unibe.ch/mailman/listinfo/moose-dev |
In reply to this post by Jan Kurš
I imagine. May be one idea would be also to have different flavors of PP.
That could be gorgeous. If Synectique takes off for real (which I really hope), I'm sure that Synectique will think hard about you as a synectiquer :). Stef
_______________________________________________ Moose-dev mailing list [hidden email] https://www.iam.unibe.ch/mailman/listinfo/moose-dev |
In reply to this post by Thierry Goubier
> > And thanks too Usman, for pointing out long source code files may (or > do) show diverging asymptotic behaviors. And real life use for most > programming languages are very long files :( > Like one method with 15 pages ? :) Stef _______________________________________________ Moose-dev mailing list [hidden email] https://www.iam.unibe.ch/mailman/listinfo/moose-dev |
In reply to this post by Tudor Girba-2
Hi Doru,
the low hanging fruit is the time needed to create a PP parser (time for new). I profiled PPSmalltalkParser over a bench used by Lucas Renggli a few years ago, and PPSmalltalkParser spends half the time in "new" (and it is a lot slower than it used to be). Thierry Le 21/02/2015 21:13, Tudor Girba a écrit : > Hi, > > On Sat, Feb 21, 2015 at 8:07 PM, Thierry Goubier > <[hidden email] <mailto:[hidden email]>> wrote: > > Le 21/02/2015 11:14, Jan Kurš a écrit : > > Hey, > > I was not aware about some exponential element in the overall > complexity, it is strange and I will investigate it. Thank you > Usman, > for pointing this out. > > In general, it is hard to add features into the PetitParser and > preserve > performance. Therefore, we work on a tool, that will take > PetitParser, > analyze it and generates faster parser. The idea is to use > PetitParser > while developing the grammar and than generate a fast parser for > "real" > use. We hope we will be present the tool during the next ESUG. > > > I certainly be interested with what you come up with. From profiling > PetitParser, it seems there is at least one low-hanging fruit in > term of performance. > > > What is that log hanging fruit? > > Doru _______________________________________________ Moose-dev mailing list [hidden email] https://www.iam.unibe.ch/mailman/listinfo/moose-dev |
In reply to this post by stepharo
Le 21/02/2015 22:04, stepharo a écrit :
> >> >> And thanks too Usman, for pointing out long source code files may (or >> do) show diverging asymptotic behaviors. And real life use for most >> programming languages are very long files :( >> > Like one method with 15 pages ? :) Only 15 pages? I believe SmaCC can generate longuer methods than that :) Thierry _______________________________________________ Moose-dev mailing list [hidden email] https://www.iam.unibe.ch/mailman/listinfo/moose-dev |
In reply to this post by Thierry Goubier
Hi,
On Sat, Feb 21, 2015 at 11:03 PM, Thierry Goubier <[hidden email]> wrote: Hi Doru, I still do not see the low hanging fruit. Do you have a concrete idea of how to make it faster? I profiled PPSmalltalkParser over a bench used by Lucas Renggli a few years ago, and PPSmalltalkParser spends half the time in "new" (and it is a lot slower than it used to be). What does "a lot slower" mean? Doru Thierry _______________________________________________ Moose-dev mailing list [hidden email] https://www.iam.unibe.ch/mailman/listinfo/moose-dev |
In reply to this post by stepharo
---> Save our in-boxes! http://emailcharter.org <---
Johan Fabry - http://pleiad.cl/~jfabry PLEIAD lab - Computer Science Department (DCC) - University of Chile _______________________________________________ Moose-dev mailing list [hidden email] https://www.iam.unibe.ch/mailman/listinfo/moose-dev |
In reply to this post by Thierry Goubier
So the easy solution in ‘production use’ would be to create the parser once and then keep hanging on to it? I would not say that this is really a PetitParser issue, but more how the parser is used. > On Feb 21, 2015, at 19:03, Thierry Goubier <[hidden email]> wrote: > > Hi Doru, > > the low hanging fruit is the time needed to create a PP parser (time for new). > > I profiled PPSmalltalkParser over a bench used by Lucas Renggli a few years ago, and PPSmalltalkParser spends half the time in "new" (and it is a lot slower than it used to be). > > Thierry ---> Save our in-boxes! http://emailcharter.org <--- Johan Fabry - http://pleiad.cl/~jfabry PLEIAD lab - Computer Science Department (DCC) - University of Chile _______________________________________________ Moose-dev mailing list [hidden email] https://www.iam.unibe.ch/mailman/listinfo/moose-dev |
Le 21/02/2015 23:24, Johan Fabry a écrit :
> > So the easy solution in ‘production use’ would be to create the > parser once and then keep hanging on to it? I would not say that this > is really a PetitParser issue, but more how the parser is used. All other Smalltalk parsers are instantiated as many times as the PetitParser one in this benchmark, and their instanciation time is around 0 ms. But, yes, I would think that one of the solution would be a pre-instantiation of a PetitParser, and a final instantiation to create an instance. In a way, one could imagine writing SmaCC parsers in the same way that PetitParser parsers are written... pre-instantiation being the part where SmaCC generates and compile its parsers. In this benchmark, this would cut parsing time by 2. Thierry _______________________________________________ Moose-dev mailing list [hidden email] https://www.iam.unibe.ch/mailman/listinfo/moose-dev |
Free forum by Nabble | Edit this page |