NUMA-Aware Multicore Matrix Multiplication

A novel user-level scheduling, along with a specific data alignment method is presented for matrix multiplication in cache-coherent Non-Uniform Memory Access (ccNUMA) architectures. Addressing the data locality problem that occurs in such systems alleviates memory bottlenecks in problems with large input data sets. It is shown experimentally that a large number of cache misses occur when using an agnostic thread scheduler (such as OpenMP 3.1) with its own data placement on a ccNUMA machine. The problem is alleviated using the proposed technique for tuning an existing matrix multiplication implementation found in the BLAS library. The data alignment with its associated scheduling reduces the number of cache-misses by 67% and consequently the computation time by up to 22%. The evaluating metric is a relationship between the number of cache-misses and the gained speedup.