[Numpy-discussion] Improving Python+MPI import performance
Dag Sverre Seljebotn
d.s.seljebotn at astro.uio.no
Fri Jan 13 16:58:56 EST 2012
On 01/13/2012 10:20 PM, Langton, Asher wrote:
> On 1/13/12 12:38 PM, Sturla Molden wrote:
>> Den 13.01.2012 21:21, skrev Dag Sverre Seljebotn:
>>> Another idea: Given your diagnostics, wouldn't dumping the output of
>>> "find" of every path in sys.path to a single text file work well?
>> It probably would, and would also be less prone to synchronization
>> problems than using an MPI broadcast. Another possibility would be to
>> use a bsddb (or sqlite?) file as a persistent dict for caching the
>> output of imp.find_module.
> We tested something along those lines. Tim Kadich, a summer student at
> LLNL, wrote a module that went through the path and built up a dict of
> module->location mappings for a subset of module types. My recollection is
> that it worked well, and as you note, it didn't have the synchronization
> issues that MPI_Import has. We didn't fully implement it, since to handle
> complicated packages correctly, it looked like we'd either have to
> re-implement a lot of the internal Python import code or modify the
> interpreter itself. I don't think that MPI_Import is ultimately the
> "right" solution, but it shows how easily we can reap significant gains.
> Two better approaches that come to mind are:
It's actually not too difficult to do something like
LD_PRELOAD=myhack.so python something.py
and have myhack.so intercept the filesystem calls Python makes (to libc)
and do whatever it wants. That's a solution that doesn't interfer with
how Python does its imports at all, it simply changes how Python
perceives the world around it ("emulation", though much, much lighter).
It does require some low-level C code, but there are several examples on
the net. I know Ondrej Certik just implemented something similar.
Note, I'm just brainstorming here and recording possible (and perhaps
impossible) ideas in this thread -- the solution you have found is
indeed a great step forward!
> 1) Fixing this bottleneck at the interpreter level (pre-computing and
> caching the locations)
> 2) More generally, dealing with this as well as other library-loading
> issues at the system level, perhaps by putting a small disk near a node or
> small collection of nodes, along with a command to push (broadcast) some
> portions of the filesystem to these (more-)local disks. Basically, the
> idea would be to let the user specify those directories or objects that
> will be accessed by most of the processes and treated as read-only so that
> those objects can be cached near the node.
> NumPy-Discussion mailing list
> NumPy-Discussion at scipy.org
More information about the NumPy-Discussion