Identify Performance Bottlenecks

    

The formulas mentioned in section 5.3.3, in addition to providing an estimate of performance, can help one identify whether the performance is limited by computation, by the number of messages, or by the volume of communication. Even if the estimate is far from correct, the user may get some information about the performance bottleneck by studying the computation and communication estimates provided by those formulas.

Comparing the execution times of a problem of size N and one of size N/2 may also provide insight into the performance of the ScaLAPACK routine being used. Let and be the time required for a problem of size N and size N/2, respectively, on P processors.

• If , the physical memory of each node may be exceeded.
• If , the performance may be limited by the rate at which flops are performed. If the flop rate is significantly less than expected, the user should check the data distribution (try the standard data distribution suggested in section 5.1.1) and the underlying BLAS.
• If , the major performance factor may be bandwidth (). This is what one should obtain for medium values of N.
• If , the major performance factor may be latency (). This is what one should obtain for small values of N.

This performance analysis suggests which computer characteristic is most likely limiting the performance. It cannot say whether one is getting good performance.