optimize KlassInfoTable size to power of 2

[臧琳]

Our preliminary measurement show the patch introduced about ~5% speed up
for jmap ¨Chisto when iterate ~13GB committed heap. ( with ¨CXX:+UseG1GC ¨CXmx180g configured)

Just FYI.

Cheers,
Lin

From: ê°ÁÕ
Sent: Monday, December 10, 2018 2:24 PM
To: serviceability-dev at openjdk.java.net
Subject: optimize KlassInfoTable size to power of 2

Dear All,
     I am investigating JMap utility for large heap (~200GB) and found that
the current KlassInfoTable¡¯s _num_buckets size(20011) may not be optimal.
     My observation is that when ¡°jmap histo¡± iterating objects , the object¡¯s klass
   is identified and then hash idx in KlassInfoTable¡¯s buckets[] is calculated by mod
of _num_bucktes, which would issue a heavy instruction idiv on x86 platform. It
means for every object scanned, a idiv instruction is issued, which lag the performance espically when
there are large number of objects in heap. Hence if the _num_buckets can be changed to
a pow of 2, (e.g. 65536) the idiv can be replaced with a faster instruction such as shl (left bit shift),
And I have prepared a patch for this change.
  My question is that why 20011 is used now? is there any special reason? And is there
any potential problem if I change the value to 65536, or 32768?
  Thanks!

BRs,
Lin
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.openjdk.java.net/pipermail/serviceability-dev/attachments/20181210/40167a15/attachment.html>