Nick Edmonds wrote:
On Sat, Sep 12, 2009 at 9:42 PM, Nick Edmonds <ngedmond@cs.indiana.edu> wrote:
yet). There's also a very nascent project to do multi-core/highly multi-threaded BGL, starting from the BGL sources as opposed to PBGL.
Any pointers ? It would be nice if they look at using the map-reduce library (boost candidate)
It's very early stages so no, no docs/code yet, it's basically some new PhD students in my lab with me and another post-doc mentoring them. I would anticipate that map-reduce is too coarse-grained/BSP-like to capture the fine grained dependencies we need to capture to get good performance in a memory-bandwidth bound domain like sparse graphs, but I'll mention it and see if they want to write a sample algorithm with it and run some numbers. We're investigating a few different options for organizing/parallelizing the work, including starting from scratch.
It would be nice to see how the map-reduce library performs though. Do you have any performance data for using it on highly multi-threaded architectures (Sun Niagara, Cray XMT, anything with >8 cores really)?
Hi Nick, the Map/Reduce library is available in the sandbox and in the vault. The performance is as much down to the implementation of the map and reduce tasks as the framework itself. Jose has compared performance of my library with Standford University's Pheonix implementation and the performance is comparable (http://lists.boost.org/Archives/boost/2009/08/155759.php). I have tested the library on MS Windows on servers with 16-cores and performance has scaled well - if you get the input's segregated sufficiently to avoid contention. I don't yet have any benchmarks, but am looking for some suitable algorithms if you're interested in contributing. Regards -- Craig