ExcessDataVolume

From HPC Wiki
Jump to: navigation, search

Description

The pattern "Excess data volume" describes the performance issue that you get, when data is not fitting into intended memory. An example is the block size for spatial blocking which need to be adjusted based on the cache size. If you excess the size of the cache size, the data needs to be fetched from a lower cache level with higher access times.


Symptoms

A simple bandwidth model is too optimistic.


Detection

  • Low utilization of the available bandwidth (in order to get reference bandwidth, see Bandwidth saturation pattern
  • Low cache hit ratio
  • Frequent CL evicts or replacements

Recommended LIKWID groups: CACHE, DATA and MEM


Possible optimizations and/or fixes

Reduce the data that should reside in memory/cache


Applicable applications or algorithms or kernels

  • All spatially blocked kernels