Dynamic G1 barrier elision for C2 in young
erik.osterlund at lnu.se
Mon Jun 8 09:59:43 UTC 2015
Since this concerns compiler too, I decided to CC hotspot-dev.
Den 06/06/15 02:44 skrev Erik Österlund <erik.osterlund at lnu.se>:
>Making G1 run faster on GC-tuned applications that are designed to only
>rarely spill objects into old, seems like an interesting and important
>optimization goal at the moment.
>Today I tried an interesting experiment. I sample garbage during the
>sweeping phase (phase 2) of System.gc() (G1MarkSweep) that stumbles
>through garbage anyway, hoping to find classes with instances that are
>used all the time, but /never/ make it into old. Then I deoptimize these
>classes and recompile the relevant nmethods depending on the class to
>elide the G1 write barriers (in C2). If the GC eventually needs to promote
>any of these objects to old, I just deoptimize again and recompile with G1
>barriers turned back on.
>On some DaCapo benchmarks, it payed off very well for a few benchmarks
>that supposedly use many temporary objects:
>fop: -9.2% time <- this one was brutal!!
>xalan: -6.9% time
>jython: -5.9% time
>Results were measured with 40 warmup iterations, and then computed the
>average of the following 10 iterations, so 50 iterations in total. Class
>unloading was turned off (using my own patch to make -Xnoclassgc work,
>because it seems to be broken currently) and 512M heaps.
>The G1 barriers are already optimized to be faster for young objects, but
>if the GC finds out that certain types of objects /never/ get old, telling
>the compiler so allows complete elision of both the pre and post barriers
>from the code which is nice.
>Are we conceptually interested in such a solution, potentially accompanied
>with a flag like -XX:+G1DynamicallyOptimizeYoung? Thought I¹d check if I
>can get some feedback before going too far with this.
>Here is the code I used.
>Patch 1: -Xnoclassgc
>This just fixes an issue that -Xnoclassgc doesn¹t work properly using G1
>(unfortunately I have yet to get the bug system work to report it...).
>With this JVM flag, it should not do class unloading. I had to run my
>experiments without class unloading because it killed the optimized
>nmethods of the almost always dead objects I want to optimize in DaCapo,
>because DaCapo does not retain their class loaders or something.
>Patch 2: Dynamic G1 barrier elision
>This is where the interesting stuff went if anyone is interested. This is
>just a very basic prototype/concept to check if the approach seems
>interesting to you guys. You probably want to add stuff like deoptimizing
>less (only if there are fields to actually optimize/deoptimize - keep
>track of that more accurately), and to sample garbage outside of
>System.gc() - this was just a convenience for now, and being more accurate
>with which class declared a field, not the canonical class, etc.
>Any comments are welcome.
More information about the hotspot-dev