Using workstations interconnected by a LAN as a distributed parallel computer is becoming increasingly common. At the same time, parallelizing compilers are making such systems easier to program, Understanding the traffic of compiler--parallelized programs running on networks is vital for network planning and for designing quality of service interfaces and mechanisms for new networks. To provide a basis for such understanding, we measured the traffic of six dense-matrix applications written in a dialect of High Performance Fortran and compiled with the Fx parallelizing compiler. The traffic of these programs is profoundly different from typical network traffic. In particular, the programs exhibit global collective communication patterns, correlated traffic along many connections, constant burst sizes, and periodic burstiness with bandwidth dependent periodicity. The traffic of these programs can be characterized by the power spectra of their instantaneous average bandwidth. These spectra can be simplified to form analytic models to generate similar traffic.

20 pages

Return to: SCS Technical Report Collection
School of Computer Science homepage

This page maintained by reports@cs.cmu.edu