Exploiting data locality in Swift/T workflows using Hercules

Abstract

The ever-increasing power of supercomputer systems is both driving and enabling the emergence of new problem-solving methods that require the efficient execution of many concurrent and interacting tasks. Swift/T, as a description language and runtime, offers the dynamic creation and execution of workflows, varying in granularity, on high-component-count platforms. Swift/T takes advantage of the Asynchronous Dynamic Load Balancing (ADLB) library to dynamically distribute the tasks among the nodes. These tasks may share data using a parallel file system, an approach that could degrade performance as a result of interference with other applications and poor exploitation of data locality. The objective of this work is to expose and exploit data locality in Swift/T through Hercules, a distributed in-memory store based on Memcached, and to explore tradeoffs between data locality and load balance in distributed workflow executions. In this paper we present our approach to enable locality-based optimizations in Swift/T by guiding ADLB to schedule computation jobs in the nodes containing the required data. We also analyze the interaction between locality and load balance: our initial measurements based on various raw file access patterns show promising results. Moreover, we present future work based on the promising results achieved so far.