> Hi, > > I have been trying to optimize or somehow control the invocation of the Java > Script Core's Garbage Collector by changing certain values like the > highWaterMark and minBytesPerCycle of Heap.cpp. By default the value of > minBytesPerCycle is 512 KB, but when i change it to 2 MB i found some > performance gain in SunSpider. > For x86 the gain is about 300 ms > For ARM(embedded) core getting around 150 ms.
You will notice that minBytesPerCycle is already set differently on different platforms precisely because of these time-space trade-offs. The Heap::Heap constructor sets it to heapSizeForHint(), which will often return something like 16MB or 8MB. If you find that you'd prefer to have finer-grained control, feel free to submit a patch that rationalizes this code to your needs. > > Is there any side affect with the changes i have made with minBytesPerCycle > from 512KB to 2 MB? As Geoff already pointed out, you are making an instance of the GC always use a minimum of 2MB of memory in steady state even if the application is not using that much. > > Is there any other i can optimize the performance or control the invocation > of GC? I recommend Steve Blackburn's papers on GC for Java as a good starting point, as those collectors are generally regarded as having the best balance of throughput, space usage, and architectural simplicity: http://users.cecs.anu.edu.au/~steveb/downloads/pdf/mmtk-sigmetrics-2004.pdf http://users.cecs.anu.edu.au/~steveb/downloads/pdf/mmtk-icse-2004.pdf You may also want to read up on the two main styles of conservative GC (Boehm and Bartlett) as JavaScriptCore uses a combination of the two: http://www.hpl.hp.com/personal/Hans_Boehm/gc/#details and http://www.hpl.hp.com/personal/Hans_Boehm/gc/#further http://computer-refuge.org/classiccmp/ftp.digital.com-jun2004/pub/Compaq/WRL/research-reports/WRL-TR-88.2.pdf And finally consider getting the new GC book by Hosking et al: http://www.amazon.com/Garbage-Collection-Handbook-Management-Algorithms/dp/1420082795/ref=ntt_at_ep_dpt_1 -Filip _______________________________________________ webkit-dev mailing list [email protected] http://lists.webkit.org/mailman/listinfo.cgi/webkit-dev
