Many-core and PSPI: Mixing fine-grain and coarse-grain parallelism

Conclusions

We implemented PSPI migration on the Sun Niagara2 by combining coarse-grained and fine-grained prallelism. We showed that the multi-thread per core model leads to significant uplift in performance over a single thread approach. Compared to a strictly fine grain parallelism we achieved a 60% uplift. Compared to a coarse grain approach the improvement was 5X. Improved floating point/vector performance could lead to signficant uplift for this aglorithm.