Adaptive Data Placement for Distributed-memory Machines Adaptive Data Placement for Distributed-memory Machines 1

Programming distributed-memory machines requires careful placement of data on the nodes. This is because achieving eeciency requires balancing the computational load among the nodes and minimizing excess data movement between the nodes. Most current approaches to data placement require the programmer or compiler to place data initially and then possibly to move it explicitly during a computation. This paper describes a new, adaptive approach to data placement. It is implemented in the Adapt system, which takes an initial data placement, eeciently monitors how well it performs, and changes the placement whenever the monitoring indicates that a diierent placement would perform better. Using Adapt can simplify the programming of parallel systems and simplify compilers for parallel languages such as HPF. In particular, Adapt frees the programmer from having to specify data placements, and it frees the compiler from having to do often complex analysis to determine a good placement. Moreover, Adapt supports a new \variable block" placement, which is especially useful for applications with nearest-neighbor communication but an imbalanced workload. For applications in which the best data placement varies dynamically, using Adapt can lead to much better performance than using any statically determined data placement. We present the performance of Adapt on three scientiic applications that require diierent data placements.