JDK-5022198 : PS 6.3 reliability test fails with JDK 1.4.2_04-b05 - orion
  • Type: Bug
  • Component: hotspot
  • Sub-Component: gc
  • Affected Version: 1.4.2_04
  • Priority: P4
  • Status: Closed
  • Resolution: Duplicate
  • OS: solaris_9
  • CPU: sparc
  • Submitted: 2004-03-29
  • Updated: 2005-12-02
  • Resolved: 2005-12-02
Related Reports
Duplicate :  
JDK-6314461 - from-space does not relinquish space at full collections.
Relates :  
JDK-5030509 - 1.4.2_04 CMS: S1PS encounters file desrciptor exhaustion only with CMS collector
Description
When a reliability tests is run with ps 6.3 from the Orion 7a installer the test fails with repeated OOM errors. We can see from the GC log that until the first Full GC, everything seems normal.  However, once Full GCs start, they don't stop even when there is plenty of memory available.

Here is a snippet of the gc log:
11740.310: [Full GC 3136670K->355594K(3141056K), 7.0372032 secs]
11748.933: [GC 739474K->679893K(3141056K), 0.0351507 secs]
11749.248: [GC 750357K->375019K(3138176K), 0.0285423 secs]
11750.578: [GC 753131K->715728K(3138176K), 0.0351470 secs]
11750.613: [Full GC 715728K->356619K(3130624K), 5.6653474 secs]
11757.414: [GC 719627K->702352K(3130624K), 0.0338336 secs]
11757.620: [GC 737232K->381458K(3127104K), 0.2411883 secs]
11759.291: [GC 737426K->704646K(3127104K), 0.0416006 secs]
11759.333: [Full GC 704646K->367965K(3122240K), 5.6040553 secs]
11766.040: [GC 714205K->704295K(3122240K), 0.0406974 secs]
11766.314: [GC 743783K->408678K(3115904K), 0.0847126 secs]
11766.399: [Full GC 408678K->347922K(3116928K), 5.8513157 secs]
11773.461: [GC 683538K->673817K(3116928K), 0.0466765 secs]
11773.642: [GC 703897K->371971K(3114944K), 0.0472885 secs]
11773.690: [Full GC 371971K->355546K(3115648K), 5.2288339 secs]

The test was run with ParallelGC.  However using ParNewGC and ConcMarkSweepGC does not improve the situation.

The jvm team suggested a workaround, specifically to disable
Comments
EVALUATION ###@###.### 2004-03-31 The largest change in the parallel scavenge collector between 1.4.1 and 1.4.2 was that adaptive sizing is on by default. Possible work around is to turn off adaptive sizing by adding -XX:-UseAdaptiveSizePolicy to the command line. ###@###.### 2004-04-27 Turning off adaptive sizing solves the problem of frequent full GCs with the parallel scavenger. ###@###.### 2004-04-27 Portal team really wants to use the CMS collector to get low pause times. Until 1.4.2_05, however, other problems prevented CMS from working properly, so they were using the parallel scavenger as a work around. The fixes in 1.4.2_05 along with some tuning of command-line options allowed Portal testing to successfully complete using CMS, so this bug is no longer a priority. Need access to a test setup to diagnose the problem with the parallel scavenger, but the orion testing is tightly scheduled and access is not likely for several or many weeks.
11-06-2004
WORK AROUND Use JDK 1.4.1_0[6-7] ###@###.### 2004-04-01 Disable adaptive sizing of the generations with the parallel scavenger (-XX:-UseAdaptiveSizePolicy).
01-04-2004