Degree

Doctor of Philosophy (PhD)

Department

The Department of Computer Science and Engineering

Document Type

Dissertation

Abstract

Asynchronous Many-Task (AMT) runtime systems are based on the idea of dividing an algorithm into small units of work, known as tasks. The runtime system is then responsible for scheduling and executing these tasks in an efficient manner by taking into account the resources provided to it and the associated data dependencies between the tasks. One of the primary challenges faced by AMTs is managing such fine-grained parallelism and the overheads associated with creating, scheduling and executing tasks. This work develops methodologies for assessing and managing overheads associated with fine-grained task execution in HPX, our exemplar Asynchronous Many-Task runtime system. Known optimization techniques, viz. active message coalescing, task inlining and parallel loop iteration chunking are applied to HPX. Active message coalescing, where messages bound to the same destination are aggregated into a single message, is presented as a solution to minimize overheads associated with fine-grained communications. Methodologies and metrics for analyzing fine-grained communication overheads are developed. The metrics identified and implemented in this research aid in evaluating network efficiency by giving us an intrinsic view of the underlying network overhead that would be difficult to measure using conventional methods. Task inlining, a method that allows runtime systems to manage the overheads introduced by a large number of tasks by merging tasks together into one thread of execution, is presented as a technique for minimizing fine-grained task overheads. A runtime policy that dynamically decides whether to inline a task is developed and evaluated on different processor architectures. A methodology to derive a largely machine independent constant that allows controlling task granularity is developed. Finally, the machine independent constant derived in the context of task inlining is applied to chunking of parallel loop iterations, which confirms its applicability to reduce overheads, in the context of finding the optimal chunk size of the combined loop iterations.

Committee Chair

Kaiser, Hartmut

DOI

10.31390/gradschool_dissertations.5106

Share

COinS