At a high level, this makes perfect sense. The details are the killer though; this is not a simple task. We had a student over the summer working on parallelization in a shared-memory environment; he made a lot of progress but nothing ready to integrate into a release. Doing it on a cluster would be much harder as you would have to package all the needed state in messages, where m5 currently assumes that passing pointers to shared objects works fine.
Steve On 11/8/07, Tony Frank <[EMAIL PROTECTED]> wrote: > Hi, > > I would like to improve the simulation speed by running the m5 in a cluster > . > > If I distribute the simulated processors and blocks of m5 into different > physical > processors, and use MPI (such as IBM blue gene ) to make them work in > parallel, does this improve simulation speed ? > > The simulated processors are only functional / behavior unit, and only > network-on-chip block is cycle-accurate . > > Does this make sense ? > > thanks > > Tony > > Nov. 8 2007 > > If the MPI packet transaction overhead is too high > > > __________________________________________________ > Do You Yahoo!? > Tired of spam? Yahoo! Mail has the best spam protection around > http://mail.yahoo.com > _______________________________________________ > m5-users mailing list > m5-users@m5sim.org > http://m5sim.org/cgi-bin/mailman/listinfo/m5-users > _______________________________________________ m5-users mailing list m5-users@m5sim.org http://m5sim.org/cgi-bin/mailman/listinfo/m5-users
