Workload characterization is central to development of new computer architectures. The rise of the mobile-cloud paradigm has increased the diversity and rate at which applications are created thus challenging computer architects' ability to build optimized systems for them. In the past, architects have been able to examine software codes of interest (often through slow laborious manual inspection if necessary) when releases were far and few in between to derive intuition necessary to make architectural and microarchitectural discoveries. But this method does not scale to emerging applications that are literally hammered out in the hundreds by the day. Further, new languages and platforms have behaviors that are quite different from legacy codes and there is an urgent need for intuition on these applications. Without new methods to characterize emerging workloads, computer architects risk running into an intuition wall. This risk might prove calamitous if unmitigated, given the added reliance on (micro)architects to develop more energy efficient designs to compensate for the losses due to slowdowns in Dennard's scaling.
Advances in machine learning provide an opportunity to overcome the intuition wall. In the last decade there have been many major advances in machine learning on graphs motivated by need/benefits of mining behaviors in social networks and enabled by cheap commodity computing. In this project, the PIs plan to leverage these advances to discover and program new computer architectures. By viewing program execution as a graph, clustering these graphs, and mining them for similarities, the PIs plan to discover new behaviors that architects and microarchitects can use to develop new on-chip acceleration structures. The PIs also plan to study how legacy code can semi-automatically be converted to execute on the architectures with the new accelerators.
