Hiding Miss Latencies with Multithreading on the Data Diffusion Machine

Large parallel computers require techniques to tolerate the potentially large latencies of accessing remote data. Multithreading is one such technique. We extend previous studies of multithreading by investigating its use on the Data Diffusion Machine (DDM), a virtual shared memory machine in which data migrates according to its use. We use a detailed emulator to study DDM’s with up to 72 nodes, allowing the scalability of multithreading to be tested further than in other studies. The results are promising and show that the applications tested can all benefit from multithreading on the DDM. Most applications however reach the ceiling of their parallelism. We briefly discuss how the results may generalise to other architectures.