Parameterisation to tailor commodity clusters to applications

摘要

Parallel applications can be parameterised by the quotient γ_a of flop/s and data transfers between the processors, and by the machine-dependent, maximum local processor performance r_a. Clusters can be parameterised by the quotient γ_m of r_a and the per processor network communication bandwidth b_m. All those parameters are predictable. In parallel machines, the communication time is smaller than the computing time if γ_m < γ_a. A first principles chemistry application is described and parameterised. Benchmarks on the Swiss-T1 cluster machine show that the predicted inter-processor communication and computing times correspond well to the measured times. These parameterisations can now be used to tailor clusters from commodity components to the applications.