Optimizing assignment of threads to SPEs on the cell BE processor

Abstract

The Cell is a heterogeneous multicore processor that has attracted much attention in the HPC community.

Performance Impact of Task Mapping on the Cell BE Multicore Processor

“…Here we see that the mapping has a much smaller influence, except for a small number of mappings. This also links our study to the work from [12]. …”

“…However, the identity permutation has gone down to rank 101 with an average runtime that is longer than the average runtime for the rank 1 permutation by a factor of 1.97. While those mappings (called ring affinity) were considered inferior in [12], that study only used uni-directional ring communication, while we consider bi-directional ring communication, which seems more realistic.…”

“…As the default mapping of the threads to the SPEs is somewhat random [12], and not necessarilty identical on two runs of the application, we also created a variant of the application where thread T i is pinned to SPU i. In order not to rely on particular implementations of the library software, we refrained from using spe set affinity, but used the mapping of the SPEs' local stores into the global address space, and provided task IDs by having the PPE directly write the IDs into a predetermined address at the local stores.…”

“…Ainsworth et al [1] conclude that among other factors that constrain the usage of Cell's Element Interconnect Bus (EIB), such as the control structure, also the concrete communication pattern leads to performance differences of the same order. Sudheer et al [12] discuss the influence of thread-SPE mapping for large messages. Their conclusions support our intuition that random thread assignments on SPEs can impact both performance and predictability significantly.…”

“…Still, the focus of their work is on finding the right affinity for a given communication pattern among a small set of possible mappings, while we are focusing more on the performance trade-offs between all possible affinities. Furthermore, in addition to the analysis of the interconnect behavior for large packets transfers done in [12], we also experiment with small and medium packets, encountered in many real-life applications (e.g., image processing convolution filters or histogram computations), and we provide interesting insights there as well.…”