Adaptive Size Policy for Best Throughput

Java GC Tutorials - Herong's Tutorial Examples

∟Parallel Collector - "+XX:+UseParallelGC"

∟Adaptive Size Policy for Best Throughput

This section provides a tutorial on how Adaptive Size Policy is used by the Parallel collector to our best Throughput goal.

In this second test, let's ask the Parallel collector to give us the best throughput performance using these JVM options:

-XX:MaxGCPauseMillis=1000
   We don't care about latency:
   You can pause application as long as 1000 ms.

-XX:GCTimeRatio=99
   We want the best throughput:
   Give us the smallest percent of time or < 1% (or 1/(1+99)) for GC

-Xmx1200m
   We don't care about footprint:
   You can use as much as 1.2GB heap

Here is what we got:

java -XX:+UseParallelGC -XX:MaxGCPauseMillis=1000
   -XX:GCTimeRatio=99 -Xmx1200m -XX:AdaptiveSizePolicyOutputInterval=1
   -Xlog:gc=debug,gc+heap=debug,gc+ergo=trace GarbageCollection2

...
Full after scavenge average_promoted 27943460
       padded_average_promoted 31022564 free in old gen 13399408
GC(39) Heap before GC invocations=40 (full 10): PSYoungGen
        total 403968K, used 5120K [...
GC(39)   eden space 398336K, 0% used [...
GC(39)   from space 5632K, 90% used [...
GC(39)   to   space 5632K, 0% used [...
GC(39)  ParOldGen       total 156160K, used 143074K [...
GC(39)   object space 156160K, 91% used [...
...
GC(39) PSYoungGen: 5120K->0K(403968K)
GC(39) ParOldGen: 143074K->32480K(161280K)
GC(39) Pause Full (Ergonomics) 144M->31M(552M) 7.945ms

GC(39) Heap after GC invocations=40 (full 10): PSYoungGen
        total 403968K, used 0K [...
GC(39)   eden space 398336K, 0% used [...
GC(39)   from space 5632K, 0% used [...
GC(39)   to   space 5632K, 0% used [...
GC(39)  ParOldGen       total 161280K, used 32480K [...
GC(39)   object space 161280K, 20% used [...

GC(39) UseAdaptiveSizePolicy actions to meet  *** throughput goal ***
GC(39)                        GC overhead (%)
GC(39)     Young generation:       14.39	  (attempted to grow)
GC(39)     Tenured generation:      4.93	  (attempted to grow)
GC(39) Tenuring threshold: (attempted to decrease to balance GC costs) = 1

9853   578813952   544504864   34309088
...

Not too bad. The Adaptive Size Policy did an ok job:

The application pause of the last GC was 7.945ms, which was way better than our latency goal of 1000 milliseconds.
Application paused for 19.32% (or 14.39 + 4.93%) of time doing GC, which was slightly worse than our throughput goal of 1%.
552M heap was used, which was better than our footprint goal of 1200M.
39 GC were performed.