For many computational tasks, the performance bottleneck is caused by memory accesses instead of CPU time. To tackle this problem in a theoretical way, the I/O-model and the Parallel External Memory (PEM) model were introduced. The PEM model describes several parallel processors, each assigned to a private internal memory (cache) of limited capacity. Communication and the storage of input, output, and intermediate results is realised by a shared external memory (disk) which is organised in blocks. This thesis considers the parallel I/O complexity of several tasks that are based on sparse matrix computations over a semiring. We present upper and lower bounds for the multiplication of a sparse matrix with (i) several vectors, including bilinear forms, (ii) a dense matrix, (iii) another sparse matrix. The work is complemented by a consideration of the MapReduce framework in the PEM model.
«
For many computational tasks, the performance bottleneck is caused by memory accesses instead of CPU time. To tackle this problem in a theoretical way, the I/O-model and the Parallel External Memory (PEM) model were introduced. The PEM model describes several parallel processors, each assigned to a private internal memory (cache) of limited capacity. Communication and the storage of input, output, and intermediate results is realised by a shared external memory (disk) which is organised in block...
»