[jboss-jira] [JBoss JIRA] (WFCORE-1324) Suggested defaults for Metasize and Java 8
Ken Wills created WFCORE-1324:
---------------------------------
Summary: Suggested defaults for Metasize and Java 8
Key: WFCORE-1324
URL: https://issues.jboss.org/browse/WFCORE-1324
Project: WildFly Core
Issue Type: Enhancement
Reporter: Ken Wills
Assignee: Ken Wills
Since PermGen is no longer used, and has been replaced by Metasize, we probably need to
alter the initial startup values. Current WF is using:
-XX:MaxMetaspaceSize=256m
After some testing with garbage collection logging on (\-verbose:gc \-Xloggc:hcgc.log
\-XX:+PrintGCDateStamps \-XX:MetaspaceSize=XX), the GC logs were monitored for at least
one occurrence of a full GC due to Metadata threshold (example [Full GC (Metadata GC
Threshold) 39592K->20187K). Using this information, minimum levels of Metasize for
various configurations were determined.
The numbers below are the values used for \-XX:MetaspaceSize=XXM followed by the number of
full GCs triggered at that amount measured during boot of WF10-full:
Standalone:
{quote}
standalone.xml 52MB(1), 53MB(0)
standalone-full.xml 64MB(1), 65MB(0)
standalone-ha.xml 52MB(1), 54MB(0)
standalone-full-ha.xml 79MB(1), 80MB(0)
{quote}
For domain mode, the corresponding values were determined to be:
{quote}
Process Controller: 12MB(1), 13MB(0)
Host Controller: 39MB(1), 40MB(0)
{quote}
In domain mode, a very slight, non-scientifically measured boot time difference was
observed (1769ms default Metasize vs 1694ms with 40m MetaSize set for host controller).
The approximate cost of increasing MetaSize over the default is summerized below (using
top to collect RSS after server boot):
JBoss AS 7.1.1: (default permgen (-XX:PermSize=256m -XX:MaxPermSize=256m), JDK 7)
||Configuration||RSS(KB)||
|standalone.xml| 182,652 |
|standalone-ha.xml | 211,672 |
|standalone-full.xml | 217,636 |
|standalone-full-ha.xml | 289,524 |
|domain:||
| Server-one: | 227,220 |
| Server-two: | 234,944 |
| PC: | 37,584 |
| HC: | 138,428 |
Wildfly 10: (default Metasize == 21M)
||Configuration||RSS(KB)||
|standalone.xml | 293,576 |
|standalone-ha.xml | 303,344 |
|standalone-full.xml | 388,660 |
|standalone-full-ha.xml | 478,576 |
|domain: ||
| Server-one: | 379,076 |
| Server-two: | 377,516 |
| PC: | 55,000 |
| HC: | 272,120 |
Wildfly 10: (Metasize == 64M)
|standalone.xml | 290,236 |
|standalone-ha.xml | 306,032 |
|standalone-full.xml | 396,596 |
|standalone-full-ha.xml | 501,576 |
|domain:|
| Server-one: |
| Server-two: |
| PC: |
| HC: |
Wildfly 10: (Metasize == 96M)
||Configuration||RSS(KB)||
|standalone.xml |317,996 |
|standalone-ha.xml | 306,516 |
|standalone-full.xml |416,008 |
|standalone-full-ha.xml |460,952 |
|domain: |
| Server-one: | 380,816 |
| Server-two: | 374,300 |
| PC: | 55,308 |
| HC: | 273,220 |
Additional measurements. Using just Wildfly-core, the following RSS sizes are measured for
the indicated Metasize:
Wildfly-10 Core master
||Metasize || RSS(KB) ||
|21m | 117,760|
|64m | 120,772|
|96m | 131,104|
There is little boot time impact on the change:
Wildfly-10 Core master
||Metasize || Boot time (MS) ||
|21m | 2127 |
|64m | 2066 |
|96m | 2099 |
Based on the memory impact of defaulting to 96M (approx 30mb initially over the default
value of 21mb), it would seem to make sense to use this as a default value, which allows
maintaining boot times without incurring a full GC due to Metasize and provides enough
initial Metasize to both start the application server and perhaps deploy an application
without incurring any performance penalty.
An additional note: host*.xml has JVM params set to MetaspaceSize=256m, which is probably
too large an initial value.
--
This message was sent by Atlassian JIRA
(v6.4.11#64026)