Why is JVM continously doing full gc while the Old Generation is only half full?

Question

I am using jdk 1.7.0_09 on a CentOS 64-bit linux machine.

The gc related vm args is

-Xmx4g -Xmn2g -XX:SurvivorRatio=4 -XX:PermSize=128m -XX:MaxPermSize=128m -XX:InitialTenuringThreshold=15 -XX:CMSWaitDuration=50 -XX:MaxTenuringThreshold=15 -XX:+UseConcMarkSweepGC -XX:+UseParNewGC -XX:CMSInitiatingOccupancyFraction=80 -XX:+CMSParallelRemarkEnabled -XX:ReservedCodeCacheSize=128m

But it's continuously doing full gc

jstat -gcutil pid 1000    
    S0     S1     E      O      P     YGC     YGCT    FGC    FGCT     GCT   
    0.00   0.00  15.62  57.52  19.90     11    2.081 11109 7454.686 7456.767   
    0.00   0.00  15.81  57.52  19.90     11    2.081 11109 7454.686 7456.767   
    0.00   0.00  15.81  57.51  19.90     11    2.081 11111 7454.892 7456.973   
    0.00   0.00  16.06  57.51  19.90     11    2.081 11111 7455.115 7457.196
    0.00   0.00  16.06  57.51  19.90     11    2.081 11111 7455.115 7457.196
    0.00   0.00  16.27  57.51  19.90     11    2.081 11111 7455.115 7457.196
    0.00   0.00  16.27  57.51  19.90     11    2.081 11111 7455.115 7457.196
    0.00   0.00  16.29  57.51  19.90     11    2.081 11111 7455.115 7457.196
    0.00   0.00  16.29  57.51  19.90     11    2.081 11111 7455.115 7457.196
    0.00   0.00  16.29  57.47  19.90     11    2.081 11113 7455.549 7457.629
    0.00   0.00  16.29  57.47  19.90     11    2.081 11113 7455.549 7457.629
    0.00   0.00  16.34  57.47  19.90     11    2.081 11113 7455.549 7457.629
    0.00   0.00  16.34  57.47  19.90     11    2.081 11113 7455.549 7457.629
    0.00   0.00  16.34  57.47  19.90     11    2.081 11113 7455.549 7457.629
    0.00   0.00  16.42  57.47  19.90     11    2.081 11114 7455.549 7457.629
    0.00   0.00  16.42  57.44  19.90     11    2.081 11115 7455.986 7458.067

score 3 · Accepted Answer · answered Nov 26 '12 at 09:10

3

Despite the column "FGC", it is not really a full GC. What the CMS is doing is a concurrent collection to see if anything can be cleaned up from the old collection (which you can see it is doing). The way to can tell this not a full GC, is that a full GC clears the eden space and copies from one survivor space to another.

IMHO The more options you use, the more likely you will come across some combination with surprising results. I would stick to the minimum of options and see how that performs. Only add option when you know they will always help.

Try to start with,

-mx4g -XX:MaxPermSize=128m -XX:+UseConcMarkSweepGC

answered Nov 26 '12 at 09:10

Peter Lawrey

525,659
79
751
1,130

I agree. The flags used indicate that this is a googled config. if you really want "-XX:InitialTenuringThreshold=15 -XX:CMSWaitDuration=50 -XX:MaxTenuringThreshold=15" you should know why. (and then you would know that they don't matter much) – Fabian Lange Nov 26 '12 at 09:12
1

What makes it worse is often the flags are purely best effort, are already enabled by default or don't do what they appears to do. Even `-ms` and `-mx` (commonly know as `-Xms` and `-Xmx`) don't do what most people think. – Peter Lawrey Nov 26 '12 at 09:17
@FabianLange We benchmarked our application against different settings of gc flags and finally settled on this. We know some config is just a suggestion. – Timmy Nov 26 '12 at 09:25

score 0 · Answer 2 · answered Nov 26 '12 at 09:11

That is normal. It you would do "-verbose:gc" you could see that the "Full" Collection is actually a concurrent one (you specified "-XX:+UseConcMarkSweepGC") You can also see that the time is not increasing by much. And that time is concurrent and not stop-the-world.

Why is JVM continously doing full gc while the Old Generation is only half full?

2 Answers2