[Gc] Problems with GC performance using gcj
bkeppler at tridentms.com
Thu Jan 13 11:20:04 PST 2011
A large portion of our churn occurs as a result of network packet traffic, so lots of byte arrays, strings, etc. that don't, in fact, need to be scanned. The interesting thing is that we went through a several day thrash to create pools of network packets to reduce heap thrash. It made a difference when running under the JVM, but not when compiled to native code with GCJ (implying the use of the Boehm GC). It appears GCJ uses GC_malloc_atomic() at the very least for Strings. I didn't do a complete analysis to prove it does so for byte arrays as well, but one assumes that since they were aware of the issue for Strings, they probably were for byte arrays as well.
One last factor to consider is that gcj appears to be using version 6.6 of the Boehm GC. I don't know how substantial the changes are in the more recent versions or whether they could have an impact on the pause times. Perhaps you could shed light on that.
Thank you for your help.
Ben Keppler, Software Engineer
Trident Micro Systems
E-mail: bkeppler at tridentms.com * Voice: 828.684.7474 * Fax: 828.684.7874
From: bruce.hoult at gmail.com [mailto:bruce.hoult at gmail.com] On Behalf Of Bruce Hoult
Sent: Tuesday, January 11, 2011 8:19 PM
To: Ben Keppler
Cc: gc at linux.hpl.hp.com
Subject: Re: [Gc] Problems with GC performance using gcj
On Sat, Jan 8, 2011 at 4:16 AM, Ben Keppler <bkeppler at tridentms.com> wrote:
Alternatively, are there settings on the Boehm GC that might relieve our problems? I would appreciate any information you could provide.
You also don't say what settings you're using now, or give any hint about the characteristics of your application.
Things that can often make a big difference to overall throughput (but not pause times) at the expense of heap size include:
GC_use_entire_heap = 1;
GC_free_space_divisor = 2;
Things that can make a big difference to pause times include GC_all_interior_pointers = 0 and using GC_malloc_atomic() instead of GC_malloc() as much as possible, especially if a large fraction of your live data is strings, sounds, images or the like that don't need to be scanned for pointers.
I don't know whether GCJ is using GC_malloc_atomic() for String, StringBuffer, byte, short, int. long, float, double but it certainly should be.
More information about the Gc