next up previous
Next: Implementation in a distributed Up: Parallel execution Previous: Results of the maximum

Efficiency of parallel computation in DeepForest

Using 20 CPUs, about 11 times speed-up was obtained. Interestingly, when we used larger data (10 MRF family's amino acid sequence data), the speed-up effect was about 17 times using 20 CPUs. It suggests DeepForest is robust to large data. However, absolute performance is not enough for actual data analyses. More sophisticated algorithm is needed. Furthermore, pragma [Chikayama et al., 1988] was not so adjusted for appropriate data stream. It might make the speed-up effect unstable (Figure 1.6).

Over 20 CPUs, speed-up effect seems to be saturated. It is probably because Cray CS-6400 is a shared memory system. We examined speed-up effect using a simple program in which little communication cost is required. Even in this case, the speed-up effect was saturated around 20 CPUs. In that sense, distributed system like PVM [Geist et al., 1998] may be more hopeful for this kind of complicated inference, even if its communication cost is higher. However, if core algorithm of DeepForest is improved, our parallel environment will be sufficient to start practical data analyses.



 

Satoshi OOta
1999-03-06