On Predicting the Impact of Resource Redistributions in Streaming Applications
Merijn Verstraaten, Sven‐Bodo Scholz · 2014
We propose a method for black-box performance modelling of executions of data-parallel array operations on shared memory multi-core systems. Black-box performance modelling refers to the idea that the source code as well as its attendant compilation process are completely independent from the modelling itself. The performance model exclusively builds on observable behaviour available when executing the compiled code. From given input characteristics and previous runtime observations we predict overall runtimes in relation to the number of cores that can be exclusively used for the task. We show that using our technique our model's runtime predictions fall within 10% of the observed runtime.