3

Having recently attempted to increase the heap space and ratio of new to old gen size I'm seeing confusing results from jstat -gccapacity which is showing much smaller capacities than I would expect.

The JVM (1.5.0_16) is started with -server -Xms2048m -Xmx2048m -XX:NewRatio=2. It's running on a Solaris 5.10 amd64 host. With about 10GB of free memory. So from what I've read the JVM should be able to make use of the full 2GB of heap space.

Watching jstat -gcutil I've observed all of the generations filling up several times causing garbage collections. E.g.:

Timestamp         S0     S1     E      O      P     YGC     YGCT    FGC    FGCT     GCT
        66150.4   0.00  51.09  56.95  90.33  54.85   6291   58.922     7   22.826   81.748

Which I would have thought would cause the JVM to expand all the generations to their full size. However, jstat -gccapacity produces:

 NGCMN    NGCMX     NGC     S0C   S1C       EC      OGCMN      OGCMX       OGC         OC      PGCMN    PGCMX     PGC       PC     YGC    FGC
700416.0 700416.0  86016.0 1408.0 1408.0  39104.0  1398784.0  1398784.0  1398784.0  1398784.0  16384.0  65536.0  38912.0  38912.0   6338     7

Subsequent runs show NGC/S0C/S1C/EC change in value:

 NGCMN    NGCMX     NGC     S0C   S1C       EC      OGCMN      OGCMX       OGC         OC      PGCMN    PGCMX     PGC       PC     YGC    FGC
700416.0 700416.0  86016.0 1472.0 1472.0  39104.0  1398784.0  1398784.0  1398784.0  1398784.0  16384.0  65536.0  38912.0  38912.0   6380     7
700416.0 700416.0 106496.0 1792.0 1856.0  97024.0  1398784.0  1398784.0  1398784.0  1398784.0  16384.0  65536.0  38912.0  38912.0   6433     7
700416.0 700416.0 106496.0 1792.0 1792.0  96064.0  1398784.0  1398784.0  1398784.0  1398784.0  16384.0  65536.0  38912.0  38912.0   6436     7

From what I understand, the capacity figures are the total size of the generation and the utilisation figures show allocations within the generation. So the above result is telling me that the combined new gen min and max capacities are both (NGCMN/MGCMX) 684MB and the old gen min and max are 1,366MB (OGCMN/OGCMX). What's confusing me is the capacities of the new generations. So my questions:

  1. Why doesn't EC + S0C + S1C == NGC? (41,920 != 86016)
  2. Why is NGC significantly smaller than NGCMN/MGCMX?
    • Is this because the max heap size is being hit (which would put it at 1,488MB, from NGC+OGC+PGC), if so what would cause the lower limit? All the documentation I can find says a Solaris 64 bit JVM should be able to use 4GB.
  3. If the max heap size is being hit why do the new generations change in capacity (either all increasing or all decreasing) without the older generation changing to compensate).

Other potentially useful jstat results:

$ jstat -gcnew 20167
 S0C    S1C    S0U    S1U   TT MTT  DSS      EC       EU     YGC     YGCT
1536.0 1600.0 1300.3    0.0 13  15 1600.0  82880.0  30892.6   6482   60.540

$ jstat -gcnewcapacity 20167
  NGCMN      NGCMX       NGC      S0CMX     S0C     S1CMX     S1C       ECMX        EC      YGC   FGC
  700416.0   700416.0   106496.0   1472.0 233472.0 233472.0   1408.0   700288.0    81088.0  6489     7

$ jstat -gcold 20167
   PC       PU        OC          OU       YGC    FGC    FGCT     GCT
 38912.0  21426.5   1398784.0   1375651.6   6503     7   22.826   83.627

$ jstat -gcoldcapacity 20167
   OGCMN       OGCMX        OGC         OC       YGC   FGC    FGCT     GCT
  1398784.0   1398784.0   1398784.0   1398784.0  6517     7   22.826   83.779

$ jstat -gcpermcapacity 20167
  PGCMN      PGCMX       PGC         PC      YGC   FGC    FGCT     GCT
   16384.0    65536.0    38912.0    38912.0  6531     7   22.826   83.925
SimonC
  • 6,590
  • 1
  • 23
  • 40

1 Answers1

2

It turns out that this is intentional behaviour by the throughput collector:

Tuning Garbage Collection with the 5.0 Java[tm] Virtual Machine > 5.2.2.2 Adjusting Generation Sizes

The statistics (e.g., average pause time) kept by the collector are updated at the end of a collection. The tests to determine if the goals have been met are then made and any needed adjustments to the size of a generation is made.

Observing the GC capacities while the JVM is under peak load does show the combined young gen current capacities to be equal to the min/max capacity values.

SimonC
  • 6,590
  • 1
  • 23
  • 40