Humongous Allocations

Parse Java G1 gc.log files that are written with -XX:+PrintAdaptiveSizePolicy enabled in the JVM options and identify humongous allocations. Humongous allocations are any allocation that exceeds half of the G1 region size.

Build (optional)

pyinstaller -F humongous_alloc.py

Usage

humongous_alloc -h
usage: humongous_alloc [-h] file [file ...]

Find humongous allocations in GC log files. JVM option -XX:+PrintAdaptiveSizePolicy is required.

positional arguments:
  file        gc.log file

optional arguments:
  -h, --help  show this help message and exit

Example

Heap Size | G1 Region | 50% of G1 Region | Allocations <= 50% G1 Region
------------------------------------------------------------------------
<4GB        1MB         0.5MB              42
<8GB        2MB         1MB                24
<16GB       4MB         2MB                61
<32GB       8MB         4MB                371
<64GB       16MB        8MB                202224
64GB+       32MB        16MB               70685

Number of Allocations >16MB: 8648

Allocation Percentiles (bytes):
	p50 = 7459136.0
	p75 = 11626040.0
	p90 = 14918248.0
	p99 = 22377104.0
Min Allocation Size: 664 bytes
Max Allocation Size: 22377368 bytes
Total Allocation Count: 282055

• The table summarizes how many allocations fit within 50% or less of a G1 region size. The purpose of this is to determine whether increasing the G1 region size would help lessen the number of humongous allocations.
  • In this example, if we had a JVM running with a 31 GB heap (-Xms and -Xmx = 31g), we have 202,224 humongous allocations we could get rid of by increasing the G1 region size to 16 MB.
  • Increasing the G1 region size can be done by:
    • Increasing -Xms and -Xmx to between 32 GB and 64 GB.
    • Increase -XX:G1HeapRegionSize to 16 MB. However, this approach should be used with caution as you will have fewer G1 regions and that can cause JVM performance issues.
• Percentile statistics are also displayed to identify what portion of the allocation are n bytes or larger.