OProfile memory mapping improved for huge pages using Java JIT code
Previously, when profiling performance of Java just-in-time (JIT) compiled code with static huge pages enabled, OProfile's operf command recorded a large number of event samples to anonymous memory (in anon_hugepage) instead of to the appropriate Java method. Operf now recognizes the static huge page identifiers and correctly maps samples to Java methods when using statically allocated huge pages.
Description of problem:
When oprofile is used to profile JIT'd code where the java process has been started with huge page memory enabled for the process a large percentage of time is reported in anon_hugepage. Oprofile does not report to see JIT'd method calls in the perf report.
If I do not start the process with hugespages, we get the correct reporting of jvm method calls.
How reproducible:
Steps to Reproduce:
1. Using the DaCapo benchmark: http://www.dacapobench.org/
2. profile jvm using: 'operf java -XX:+UseLargePages -agentlib:jvmti_oprofile -jar dacapo-9.12-bach.jar sunflow'
3. view captured data: 'opreport --symbols --threshold 1.0'
Actual results:
samples % image name symbol name
582252 84.4164 anon_hugepage (deleted) /anon_hugepage (deleted)
68864 9.9841 libjvm.so /usr/lib/jvm/java-1.7.0-openjdk-1.7.0.71-2.5.3.1.el7_0.x86_64/jre/lib/amd64/server/libjvm.so
23543 3.4133 no-vmlinux /no-vmlinux
11087 1.6074 libjava.so /usr/lib/jvm/java-1.7.0-openjdk-1.7.0.71-2.5.3.1.el7_0.x86_64/jre/lib/amd64/libjava.so
Expected results:
samples % image name symbol name
157817 21.8364 2838.jo void org.sunflow.core.accel.KDTree.intersect(org.sunflow.core.Ray, org.sunflow.core.IntersectionState)
147963 20.4729 2838.jo Interpreter
131928 18.2542 2838.jo void org.sunflow.core.accel.BoundingIntervalHierarchy.intersect(org.sunflow.core.Ray, org.sunflow.core.IntersectionState)~1
104371 14.4413 2838.jo void org.sunflow.core.primitive.TriangleMesh$WaldTriangle.intersect(org.sunflow.core.Ray, int, org.sunflow.core.IntersectionState)
79441 10.9919 libjvm.so /usr/lib/jvm/java-1.7.0-openjdk-1.7.0.71-2.5.3.1.el7_0.x86_64/jre/lib/amd64/server/libjvm.so
37384 5.1726 2838.jo org.sunflow.image.Color org.sunflow.core.gi.InstantGI.getIrradiance(org.sunflow.core.ShadingState, org.sunflow.image.Color)~1
23981 3.3181 no-vmlinux /no-vmlinux
11256 1.5574 libjava.so /usr/lib/jvm/java-1.7.0-openjdk-1.7.0.71-2.5.3.1.el7_0.x86_64/jre/lib/amd64/libjava.so
9761 1.3506 2838.jo void org.sunflow.core.light.TriangleMeshLight$TriangleLight.getSamples(org.sunflow.core.ShadingState)
Created attachment 978413[details]
recognize the perf mmap anon_hugepage information
Something like the attached patch should recognize the slightly different information for the anon_hugepage mmap information from perf. However, this has not been tested on rhel6.
Successfully reproduced with oprofile-0.9.9-6.el6.
When tested with oprofile-0.9.9-7.el6 no issue occured.
The value 128 has been used for nr_hugepages.
Since the problem described in this bug report should be
resolved in a recent advisory, it has been closed with a
resolution of ERRATA.
For information on the advisory, and where to find the updated
files, follow the link below.
If the solution does not work for you, open a new bug report.
https://rhn.redhat.com/errata/RHBA-2015-1367.html