Re:Doom 3 source released
by anaisbetts on Tuesday November 22, 2011 @07:36PM
(#3268145)
Most game projects do end up having to write a custom allocator, due to very specific memory budgets, deterministic malloc/free perf requirements, and issues with fragmentation.
That has been borne out by my experience so far too.
There have been several successful efforts at writing better performing general-purpose allocators:
* TCMalloc (http://goog-perftools.sourceforge.net/doc/tcmalloc.html)
* jemalloc (http://www.canonware.com/jemalloc/)
There's a great Facebook Engineering blog post about jemalloc which describes why a fast allocator is needed & why the standard malloc is not always sufficient:
https://www.facebook.com/notes/facebook-engineering/scalable...
Re:Doom 3 source released
by forrestthewoods on Tuesday November 22, 2011 @08:06PM
(#3268237)
I've worked with several projects where a TCMalloc based allocator (threaded) was used to great success. A simple switch to DLMalloc (non-threaded) can also provide a significant performance boost.
Moderate usage of STL can do nasty, nasty things to memory performance. In one case a game I worked on ran at 5-10fps and a simple, fast change to DLMalloc instantly took perf to 60fps.
It's worth noting that at least on some platforms the system malloc is dlmalloc (presuming that DL is Doug Lea).
Which platforms?
glibc's malloc has long been based on Doug Lea's malloc.
+1 for TCMalloc.
We used TCMalloc at Opera Community[0], and they still might be using it there for all I know ... it greatly improved performance on our MySQL servers!
[0]: http://my.opera.com/
Aren't you using Percona build ?
Not when I worked there, but it's a few years ago. Don't know what they use now.
No, the point of a custom allocator is writing one that works with your specific requirements. A general allocator will never be more efficient than a custom one designed for the specific task. It becomes just an issue of is the cost worth what the optimization buys you, and games are an area where you need every drop of speed you can get.
That makes very little sense to do - every single project? Don't you think someone would have released a library to implement the "new and improved allocator" if everyone needs to use it?