How significant is the latency of a processor accessing the memory not directly in its local region (but still in the shared address space) compared to latency of accessing memory in a distributed address space? Is it significant enough to be worth considering seriously when working on a shared address space, or can we more or less disregard this latency?
vasua
I think you can roughly equate anything that's not in local address space as being a network communication, on the order of milliseconds rather than the 2-3 clock cycles for cache hits; in other words, avoid network communication at all costs. Efforts are obviously being made to reduce communication costs (e.g. Infiniband) in server rooms, but it's still really tough to beat the communication capabilities of being nanometers from something.
How significant is the latency of a processor accessing the memory not directly in its local region (but still in the shared address space) compared to latency of accessing memory in a distributed address space? Is it significant enough to be worth considering seriously when working on a shared address space, or can we more or less disregard this latency?
I think you can roughly equate anything that's not in local address space as being a network communication, on the order of milliseconds rather than the 2-3 clock cycles for cache hits; in other words, avoid network communication at all costs. Efforts are obviously being made to reduce communication costs (e.g. Infiniband) in server rooms, but it's still really tough to beat the communication capabilities of being nanometers from something.