In unit 2, we have gone through the use of distributed shared memory in parallel computer architecture. Although the use of distributed shared memory has the difficulty of accessing the remote memory, which results in latency troubles. This difficulty increases in case of large-scale multiprocessors like massively parallel processors (MPP).
For example, single processor in a multiprocessor system requires two memory loads of two variables by two remote processors as shown below in Figure. The issuing processor will use these variables concurrently in one operation. In case of large-scale MPP systems, the following two troubles arise:
Latency problems in MPP