Did you actually find many machines/distribs that don't have libxml2
installed by default? There are literaly hundreds of packages that
depend on libxml2 (at least in Debian) so I am not sure depending on
it is really a problem.
Also are there really some string space problems? Even when talking
about 1000 nodes transferring 100kB once at the beginning on the
job, it doesn't look too bad to me (and these XMLs could be cached
on the frontend as long as the compute nodes don't change).
Otherwise, implementing this is likely easy, especially if you find
somebody to do it :) Start from the XML export, convert it into a
text export, and write the corresponding import (starting from the
XML import may be hard because it's recursive).
Would you need an export to a file or to a memory buffer or both?
Last but not least: what's the deadline?
Le 01/09/2011 17:30, Jeff Squyres a écrit :
We're (finally) bringing full hwloc services up in Open MPI.
One of the things we want to do is send server topologies from back-end compute nodes to the front-end node. The XML export/import functionality would work for this, but a) it's a bit heavyweight, and b) it seems weird to require XML to build MPI.
Is there any chance that a lighter-weight, simple string parsing module could be added to hwloc? I'm guessing that we could save a modest amount of string space (SWAG: 20%?), but we wouldn't need a dependency on libxml, which would be good.
I took a lstopo --no-io foo.xml output on an older xeon machine and, while sitting on a boring teleconf, I manually converted it in emacs to a (slightly) simpler text format. I attached the two files. There's a modest space savings (about 17%). But libxml clearly would not be necessary.
Do you think this would be easy to implement?
hwloc-devel mailing list