Could someone tell me how these settings are used in OMPI or give any
guidance on how they should or should not be used?
The background is that (on Linux? with GCC libc? with OMPI?) small
memory allocations are allocated on the heap, with brk() or sbrk() used
to modify the high-water mark. Lest a large, freed allocation cannot be
returned to the OS due to a small, active allocation, the memory
allocator uses mmap() instead of brk/sbrk for large allocations. There
is some discussion of the internet about how mmap is a costly way of
allocating memory, but I'm concerned about something else. With mmap,
you get page-aligned allocations back. This means that if you loop over
the elements of multiple large arrays (which is common in HPC), you can
generate a lot of cache conflicts, depending on the cache associativity.
There are multiple reasons one might want to modify the behavior of the
memory allocator, including high cost of mmap calls, wanting to register
memory for faster communications, and now this cache-conflict issue.
The usual solution is
setenv MALLOC_MMAP_MAX_ 0
setenv MALLOC_TRIM_THRESHOLD_ -1
or the equivalent mallopt() calls.
This issue becomes an MPI issue for at least three reasons:
*) MPI may care about these settings due to memory registration and
pinning. (I invite you to explain to me what I mean. I'm talking over
my head here.)
*) (Related to the previous bullet), MPI performance comparisons may
reflect these effects. Specifically, in comparing performance of OMPI,
Intel MPI, Scali/Platform MPI, and MVAPICH2, some tests (such as HPCC
and SPECmpi) have shown large performance differences between the
various MPIs when, it seems, none were actually spending much time in
MPI. Rather, some MPI implementations were turning off large-malloc
mmaps and getting good performance (and sadly OMPI looked bad in
*) These settings seem to be desirable for HPC codes since they don't
do much allocation/deallocation and they do tend to have loop nests that
wade through multiple large arrays at once. For best "out of the box"
performance, a software stack should turn these settings on for HPC.
Codes don't typically identify themselves as "HPC", but some indicators
include Fortran, OpenMP, and MPI.
I don't know the full scope of the problem, but I've run into this with
at least HPCC STREAM (which shouldn't depend on MPI at all, but OMPI
looks much slower than Scali/Platform on some tests) and SPECmpi
(primarily one or two codes, though it depends also on problem size).
Discussion is invited.