Section: New Results

Cross Platform Classification for Detecting Locality Sensitivity and Selecting Data and Threads Placement Strategy

Individual nodes composing High Performance Computing (HPC) systems embed complex multicore and manycore processors. At this scale, compute tasks and data placement can double or halve execution times with respectively trivial are wise placements. While state of the art placement solutions can offer good performance improvements, they failed to set up as standards in supercomputers software stack. Current solutions are rather directed toward data or thread driven static policies. Among existing or promising future placement solutions a deep evaluation of applications response to these had yet to be done in order to wisely choose the best one.

With a set of 37 HPC representative applications, three different HPC processors, and 51 state of the art characterization metrics we built thousands models to evaluate applications response to data and threads placement policies. Thanks to a thorough methodology, our models were able to predict applications sensitivity to locality and their preferred placement policy both on new platforms and new applications. In the first case we were able to achieve more than 75% accuracy while preferred policy predictions approach optimal speedups in the second case.

This work was conducted using the PlaFRIM experimental testbed, in collaboration with Thomas Ropars from Laboratoire d'Informatique de Grenoble.

Several leads can be taken toward an extension of this work. For instance, predictions can be improved with benchmark directed learning. Models interpretation can also be furthered studied to refine the design of application characterization metrics.