A short description of the setup (w/ or w/o InfiniBand, total amount of RAM, etc) and its performance for matrix vector multiplication (similar to the picture below) would be highly appreciated.
Note that the matrix size has to be big enough to trigger network data transfer.
Ultimately I am interested in KVM (OpenStack) overhead in that situation (so I guess similar experience with other clouds should also be very informative).
PS also look at a nice answer explaining ScaleMP "features" http://serverfault.com/a/214185/106906
What is actually proposed by vSMP guys is to run cloud on top of ScaleMP.. But I do not know if they already support OpenStack or compatible cloud standards? It makes an impression of partially overlapping functionality..and I wonder if this is only "marketing" or indeed something that is worth looking at.
In any case, for now I am curious only about transparent RAM-over-network mapping technology. Does OpenStack already provide an alternative to this?