Next: Non-uniform memory access machines

Up: Departmental seminar at UWCC

Previous: Problems addressed by ScaLAPACK

ScaLAPACK Key Ideas

• Use block-partitioned algorithms to maximize data reuse in lower levels of memory
  • reduce cache misses
  • reduce frequency of communication
• Use Parallel BLAS (PBLAS) as main computational building blocks.
• Use Basic Linear Algebra Communication Subprograms (BLACS) to perform communication
• Hide parallelism within the PBLAS
• Matrices and vectors are global objects
• Fine-tune performance by adjusting data layout parameters

Important: The PBLAS make use of the sequential BLAS for which assembly coded versions exist for many processors.