Near-optimal Deployment of Dataflow Applications on Many-core Platforms with Real-time Guarantees

Safe and optimal deployment of data-streaming applications on many-core platforms requires the realistic estimation of task Worst-Case Execution Time (WCET). On the other hand, task WCET depends on the deployment solution, due to the varying number of interferences on shared resources, thus introducing a cyclic dependency. Moreover, WCET is still an over-approximation of the Actual Execution Time (AET), thus leaving room for run-time optimisation. In this paper we introduce an offline/online optimisation approach. In the offline phase, we first break the cyclic dependency and acquire safe and near-optimal solutions for tasks partitioning/placement, mapping, scheduling and buffer allocation. Then, we tighten the WCETs and update the scheduling function accordingly. In the online phase we introduce a safe distributed readjustment of the offline schedule, based on the AET. Experiments on a Kalray MPPA-256 platform show a tightening of the guaranteed latency up to 46% in the offline phase, and 41% latency reduction in the online phase. In total, we achieve more than 50% of latency reduction.

Published in:
Proceedings Of The 2017 Design, Automation & Test In Europe Conference & Exhibition (Date), 752-757
Presented at:
20th Conference and Exhibition on Design, Automation and Test in Europe (DATE), EPFL Campus, Lausanne, SWITZERLAND, MAR 27-31, 2017
New York, IEEE

 Record created 2017-09-05, last modified 2018-10-30

Rate this document:

Rate this document:
(Not yet reviewed)