Re: how is malloc/free done in SMP linux

Re: how is malloc/free done in SMP linux

[H.J. Lu]

> 
> Sir;
> 
> i am beginning development of a CAD like program. it will be
> threaded and multi-processor capable (well thanks to SMP O/S
> like linux).
> 
> as to the linux's memory management, suppose a program has
> 30 threads. each thread is allocating and deallocating memory.
> 
> does linux block the other 29 threads when thread <X> wants
> to malloc/free? how about other threads/processes in the process
> table? 
> 
> in a regular, single-processor system, my original C++ program
> spent 50% of its time in malloc/free. i reduced this time very
> significantly by implementing memory pools. the basic algorithm is fairly
> easy to split up across threads. hence linux SMP for more performance.
> but i am wondering if i might be walking into a "technical snare"
> with respect to linux and memory management.
> 

This has little to do with kernel. I forwarded this to the glibc
mailing list.

-- 
H.J. Lu (hjl@gnu.org)

Re: how is malloc/free done in SMP linux

[Wolfram Gloger]

> > as to the linux's memory management, suppose a program has
> > 30 threads. each thread is allocating and deallocating memory.
> > 
> > does linux block the other 29 threads when thread <X> wants
> > to malloc/free?

Usually, no.  Only when two (or more) threads want to simultaneously
free() chunks from _one_ particular arena, all those threads except
for one will be blocked.  In practice, this is very rare.

malloc() will actually _never_ block unless more memory is needed from
the kernel and the sbrk() or mmap() system call blocks.

> how about other threads/processes in the process
> > table? 

Not quite sure, but generally I would expect them to be unaffected by
a malloc() or free() in a particular process, unless there is (heavy)
swapping activity going on.

> > in a regular, single-processor system, my original C++ program
> > spent 50% of its time in malloc/free. i reduced this time very
> > significantly by implementing memory pools. the basic algorithm is fairly
> > easy to split up across threads. hence linux SMP for more performance.
> > but i am wondering if i might be walking into a "technical snare"
> > with respect to linux and memory management.

Linux's malloc may perform well enough so you don't need to implement
pools yourself for SMP.

Regards,
Wolfram.
-- 
`Surf the sea, not double-u three...'
wmglo@dent.med.uni-muenchen.de