[Gc] Benchmarks for 2 GB process
bbeuning at corecard.com
Fri Dec 21 08:44:08 PST 2012
Thanks for the information.
Our processes are usually around 1 GB and sometimes get to 2 GB,
but your point about memory retention still applies.
Porting a large code base to 64-bit is harder than it sounds.
We have had 64-bit running for years but we have just recently
gotten the memory usage back down close to the 32-bit process
memory usage. The 64-hit still runs a little slower than the
I tried to run a test using Boehm GC (with GC disabled and using
GC_free) as our memory allocator but it is crashing so I obviously
did something wrong.
----- Original Message -----
From: "Bruce Hoult" <bruce at hoult.org>
To: "Brian Beuning" <bbeuning at corecard.com>
Cc: "gc" <gc at linux.hpl.hp.com>
Sent: Tuesday, December 18, 2012 7:05:28 PM
Subject: Re: [Gc] Benchmarks for 2 GB process
It's important to distinguish between:
- 2 GB of process size that includes mostly pointer-free data such as images or sounds or arrays of scientific data (which should be allocated using GC_malloc_atomic), and a much smaller amount of memory that actually needs to be scanned in the mark phase of a GC
- 2 GB of process size that consists of pointer-full objects that all need to be scanned.
The latter would be quite unusual, though it is of course possible, and will be slow (but that might be unavoidable). The former will be high performance.
Either way, if you have 2 GB in the GC heap then you should be using 64 bit code, otherwise you'll start to have big problems with random integers looking like pointers into the heap and a lot of retention of memory that should have been freed.
Just random thoughts that don't really bear on your question as such...
On Wed, Dec 19, 2012 at 3:51 AM, Brian Beuning < bbeuning at corecard.com > wrote:
I Googled for "Boehm GC benchmarks" and read some of the results.
They all seem to be for smallish (30 to 200 MB) process sizes.
Are there any GC benchmark results for process sizes around 2 GB.
Note - I have worked with Boehm GC before. About 15 years ago
we converted it to allocate from shared memory.
Gc mailing list
Gc at linux.hpl.hp.com
More information about the Gc