T3E E-registers

• Benchlib library

  • One sided data transfers from memory to E-registers bypassing cache
  • Scatter / Gather in hardware
  • Nonblocking
  • More complicated to use than SHMEM
  • Not supported by Cray
  • 592 MB/sec vs 401 MB/sec for copy

Previous slide

Next slide

Back to first slide

View graphic version