Architectural investigation of matrix data layout on multicore processors

摘要

Many practical applications include matrix operations as essential procedures. In addition, recent studies of matrix operations rely on parallel processing to reduce any calculation delays. Because these operations are highly data intensive, many studies have investigated work distribution techniques and data access latency to accelerate algorithms. However, previous studies have not considered hardware architectural features adequately, although they greatly affect the performance of matrix operations. Thus, the present study considers the architectural characteristics that affect the performance of matrix operations on real multicore processors. We use matrix multiplication, LU decomposition, and Cholesky factorization as the test applications, which are well-known data-intensive mathematical algorithms in various fields. We argue that applications only access matrices in a particular direction, and we propose that the canonical data layout is the optimal matrix data layout compared with the block data layout. In addition, the tiling algorithm is utilized to increase the temporal data locality in multilevel caches and to balance the workload as evenly as possible in multicore environments. Our experimental results show that applications using the canonical data layout with tiling have an 8.23% faster execution time and 3.91% of last level cache miss rate compared with applications executed with the block data layout.