It would probably depend on how fast the Infiniband is. For a system of this calibur you'd gonna need at least 20 GB/s of sustained bandwidth to have any hope of being able to use it efficiently as shared memory. I'm also unsure of how the high latency is going to play out. Perhaps HyperThreading will be able to cover up most of those delays. I don't know though.
Lemme know when it's ready so I can send you a binary with the high-end algorithm threshold dropped to 1 billion (or even lower). If the performance scaling turns out to be okay on two motherboards, then you can try going higher. That NUMA-friendly algorithm is NUMA friendly because it's heavily optimized to simply not use memory until it's absolutely needed. But it isn't actually "aware" of the NUMA. By comparison, most of the algorithms thrash memory all over the place.




lol
Reply With Quote


Bookmarks