Temporal characterization of demands for data movement on parallel programs

A Metric for the Temporal Characterization of Parallel Programs

“…The difficulty is in establishing a temporal correspondence between operations that occur on distinct processors. This is done using the execution models developed in [8] for barrier and critical section synchronized programs. The distance between operations is computed as follows.…”

“…In [8] a machine-independent notion of the time of occurrence of events in a shared address space parallel program was developed. Parallel program performance depends on the timing relation of events in parallel execution, so the time of occurrence of these events must be described in a machineindependent way in order to achieve a machine-independent metric.…”

“…Analysis of a parallel program that is statically scheduled on a fixed number of processes can conclude that two operations will be separated by tens or thousands of instructions. This idea is formalized and made more precise in [8] and is reviewed in the next section. The two important problems in this process are determining the number of operations executed between two events (possibly on distinct processors) and mapping operation count to time.…”

“…All these studies also assume static knowledge of work scheduling. The results of experiments on the KSR-1 multiprocessor reported in [8] show that it is possible to statically determine the occurrence of interprocess reads.…”

“…In Section 2 we describe the IR metric proposed in [8]. We define the metric and show its behavior using simple traces.…”