Measurements, Proper Scaling and Application of Communication Parameters

The measurements of the communication parameters for different parallel computers are presented. These parameters are scaled to equivalent lost operations for startup and transfer of one operand. Then latency hiding is discussed and measured. For the example of the parallel matrix multiplication on p processors we present the theory for three different types of scaling: same problem, same storage per processor and same computation per processor. The theory is compared to measurements for the CRAY T3D, Intel Paragon, IBM RS/6000 SP with ATM-Switch and with High Performance Switch 3. Additional system overhead, compared to basic measurements, scales by a negative overlap factor.