Hi, this will be a long mail. I have organized it in different sections. Resume: ====== - There are performance issues in Pharo 7. - I have made benchmarks. - With Guille we made an improvement in saving source code in methods. - More improvements to come. - Automation and CI in the future. Background ========= Since the release of Pharo 7 (and before it), it was notorious that there exists a performance regression in the normal execution of the image. By checking the usual operations, we have seen ( and many have also detected) that there was an issue with the loading, compilation, and unloading of code. Also with the creation of classes, traits and the use of slots. Benchmarks ========= Although we were sure that there is a performance regression, we have to show it in a way we can test it and measure it. If we cannot measure it or repeat its execution it is worthless. For doing so, I have created an initial set of micro-benchmarks to cover normal operations of the image. The set of benchmarks is available here: https://github.com/tesonep/pharo-benchmarks These benchmarks are developed on top of SMark, only adding a command line tool and the ability to generate a CSV file. The idea is to run the benchmarks in different versions of Pharo an assert that we are not breaking anything. The first results were basically a nightmare, some operations take almost 20 times more in Pharo 7. Especially, the ones that are compiling methods. In the attached document, there is the detail of all the benchmarks, the different results and the analysis of the improvements and regressions (Positive percentages are regressions (more time), negative are improvements (less time)). I have checked the results in OSX with 64 bits images. But as the problem is in pure Smalltalk implementations the problems are (and the solutions) cross platforms. First Improvement ============== Having the benchmarks, it was easy to start looking for the problems. Thanks to the help of Guille we have detected some problems in the implementation of SourceFile. Objects of this class have the responsibility to handle the save of the source code when a method is modified. Improving this implementation we have gotten to results similar to Pharo 6 in the compilation of methods. Comparing a stock Pharo8 image with the one with the fix, we have the following improvements: AddingAnInstanceVariableToAClassWithASubclass -3.96% AddingAnInstanceVariableToAClassWithoutASubclassAndInstances -15.91% AddingAnInstanceVariableToAClassWithoutSubclasses -28.50% ClassAndSubclassCreation -25.67% ClassUsingATrait -90.05% SimpleClassCreation -26.55% TraitCreation -92.95% TraitCreationUsingOtherTrait -91.68% CompileAMethodAndModifyIt -32.92% CompileAMethodInAClassUsingATrait -83.28% CompileAMethodInATraitNotUsed -85.12% CompileAMethodInATraitUsed -88.16% CompileANewMethod -40.73% CompileANewMethodUsingInstanceVariable -33.93% MethodAccessingAGlobal -47.57% Again there are more details in the attached file. Also, we have ported this fix to Pharo 7. Next Steps ======== - Making it a part of the CI infrastructure: making it run in each PR and build to detect regressions in the changes. - Adding more micro and macro benchmarks. I have a list of things to test, but I am open to more: # Micro - Slot Implementation - Process handling - Refactorings - Files (open / write / read) # Macro - Loading: Moose / Seaside - Recompile All - Condense Sources - Startup - Shutdown We also know that there are platform related issues (especially Windows), so the idea it will be the same, build a benchmark, measure it, improve it. The idea is to have a more robust way of detecting and handling the performance of Pharo. Of course, I am open to all your comments. Cheers, Pablo