Performance evaluation of task grain programs
V.I. Ivanov · 2002
Behavioral programs are graph like objects that describe the execution of parallel programs supplied with given inputs. They quantify the amount of computation a run entails and outline the run time data dependencies. Other characteristics of the real machine (e.g. the multiprocessor management overhead oh, the communication delay dy, or the round robin time quanta tq) further affect performance directly, and indirectly via changes in allocation. Our experiments with the behavioral simulator show that linear variation of elapsed time with dy and oh/tq, predicted theoretically for ideal machines, holds for real machines as well; however, the complexity introduced by allocation and its dependency on all the other factors prevent a more accurate performance prediction.>