2009/7/8 erik quanstrom <quans...@coraid.com>: >> But don't underestimate the value of the interesting ideas in the >> linux kernel that get the performance, e.g. RCU. I don't think there >> are any OSes that have scaled to 4096 CPUs at this point besides >> Linux. > > i thought that massively parallel harvard-arch machines had > generally fallen out of favor in favor of blue gene-style hardware. > > is this incorrect?
I think it depends on the application. I have a friend who studies fluid dynamics for scramjets and he was mentioning how doing some of those calculations requires ultra-low latency. The algorithms they need to use require multiple passes, and each calculation is affected by surrounding `blocks.' With infiniband and whatnot, this might be moot (he's doing his stuff on 32-core systems right now, so it's not even to that degree), but perhaps there are still applications that still significantly benefit from that architecture. They are quite the opposite of `embarrassingly parallel' problems (a la distributed.net / SETI / Folding). That said, I have no idea what the performance / latency characteristics are of a system with 32 cores and 32 CPUs connected with infiniband / some other proprietary high speed interconnect, or what that would look like performance-wise if it was scaled to 4096 cores versus quad-CPU boards with interconnects. > - erik -dho
