Isometry: A Path-Based Distributed Data Transfer System

Publication image

Data transfers in parallel systems have a significant impact on the performance of applications. Most existing systems generally support only data transfers between memories with a direct hardware connection and have limited facilities for handling transformations to the data’s layout in memory. As a result, to move data between memories that are not directly connected, higher levels of the software stack must explicitly divide a multi-hop transfer into a sequence of single-hop transfers and decide how and where to perform data layout conversions if needed. This approach results in inefficiencies, as the higher levels lack enough information to plan transfers as a whole, while the lower level that does the transfer sees only the individual single-hop requests.

We present Isometry, a path-based distributed data transfer system. The Isometry path planner selects an efficient path for a transfer and submits it to the Isometry runtime, which is optimized for managing and coordinating the direct data transfers. The Isometry runtime automatically pipelines sequential direct transfers within a path and can incorporate flexible scheduling policies, such as prioritizing one transfer over another. Our evaluation shows that Isometry can speed up data transfers by up to 2.2× and reduce the completion time of high priority transfers by up to 95% compared to the baseline Realm data transfer system. We evaluate Isometry on three benchmarks and show that Isometry reduces transfer time by up to 80% and overall completion time by up to 60%.


Zhihao Jia (Stanford University)
Sean Treichler (NVIDIA)
Galen Shipman (Los Alamos National Laboratory)
Patrick McCormick (Los Alamos National Laboratory)
Alex Aiken (Stanford University)

Publication Date

Uploaded Files