Stefano Mazzocchi wrote:
have we considered using
http://goog-perftools.sourceforge.net/
This could possibly be useful for the VM's internal memory allocation,
but I doubt if there would be a noticeable win.
If I look at
http://wiki.wikked.net/wiki/Squid_memory_fragmentation_problem
it shows you some serious potential for improvement in memory management
(but I don't know enough about our GC to know how useful that is)
Thoughts?
In a garbage collected system the issues are very different. I haven't
studied xiaofeng's code, but in a standard implementation of a
generational GC with a mark-compact nursery:
- allocation is done with a bump-pointer. The actual allocation can be
done in something like 7 instructions (although I'm not sure whether in
DRLVM the fetch of the pointer from the TLS can be optimized away yet).
Appel's paper is the standard reference for this.
- Each thread grabs a chunk of memory and allocates from it without
synchronization, then uses a lock to synchronize access to the global
pool when it needs another chunk. The synchronization cost of
allocation is virtually negligible given an appropriate chunk size.
- The various malloc implementations are free-list allocators. They
need to maintain lists of free and allocated memory. They need to worry
about merging adjacent free blocks, finding a best-fit block in a list
of variable size free blocks, fragmentation and all that. In a copying
GC, we don't have that problem because we can move objects around.
The main ways to get performance out of a garbage collected memory
subsystem are:
- co-locating objects that exhibit temporal locality
- Engineering the mechanisms with minimal overhead (eg the helper
inlining work now being done)
- Choosing algorithms that minimise copying overhead. Pre-tenuring
objects we know will be long-lived. Generational collection.
In short, there has been work on improving performance of both garbage
collected and explicit memory management work, which overlap in the
design of free-list allocators (used by mark-sweep and reference counted
collectors), but beyond that the two are very different worlds.
cheers
--
Robin Garner
Dept. of Computer Science
Australian National University
http://cs.anu.edu.au/people/Robin.Garner/
