CAF Performance Specifications Across Platforms
We have implemented the core features of CAF, enabling us to express non-trivial CAF programs. Our compiler performs source-to-source transformation of CAF codes into F90 plus calls to a communication library (currently ARMCI). This strategy was designed to leverage the best back-end compiler available on the target platform to optimize local computation.
Performance specifications were measured across three platforms and can be reviewed below.
Alpha+Quadrics | Itanium 2+Myrinet 2000 | SGI Altix3000
Alpha+Quadrics Results
The efficiency of BT is low because of the sequential code performance. We are currently working on fixing this problem.
Itanium 2+Myrinet 2000 Results
SGI Altix3000 Results
The results were measured on a loaded system, with users competing for memory bandwith.