4.7 Article Proceedings Paper

On the Design of Minimal-Cost Pipeline Systems Satisfying Hard/Soft Real-Time Constraints

Journal

Publisher

IEEE-INST ELECTRICAL ELECTRONICS ENGINEERS INC
DOI: 10.1109/TETC.2017.2788800

Keywords

High-level synthesis; heterogeneous pipeline systems; probabilistic execution times; approximation algorithms

Funding

  1. National 863 Program [2015AA015304]
  2. National Natural Science Foundation of China [61472052]
  3. China Scholarship Council [201706050116, 201706050117]

Ask authors/readers for more resources

This paper addresses the design of application-specific pipelines in heterogeneous architectures, modeling execution times as random variables to overcome the challenge of uncertainty. By proving the NP-hardness of the problem, presenting Mixed Integer Linear Programming (MILP) formulations, and devising an efficient (1 + epsilon)-approximation algorithm, significant cost reductions are achieved, reaching up to 31.93% on average according to experimental results.
PPipeline systems provide high throughput for applications by overlapping the executions of tasks. In the architectures with heterogeneity, two basic issues in the design of application-specific pipelines need to be studied: what type of functional unit to execute each task, and where to place buffers. Due to the increasing complexity of applications, pipeline designs face a bundle of problems. One of the most challenging problems is the uncertainty on the execution times, which makes the deterministic techniques inapplicable. In this paper, the execution times are modeled as random variables. Given an application, our objective is to construct the optimal pipeline, such that the total cost of the resultant pipeline can be minimized while satisfying the required timing constraints with the given guaranteed probability. We first prove the NP-hardness of the problem. Then, we present Mixed Integer Linear Programming (MILP) formulations to obtain the optimal solution. Due to the high time complexity of MILP, we devise an efficient (1 + epsilon)-approximation algorithm, where the value of epsilon is less than 5 percent in practice. Experimental results show that our algorithms can achieve significant reductions in cost over the existing techniques, reaching up to 31.93 percent on average.

Authors

I am an author on this paper
Click your name to claim this paper and add it to your profile.

Reviews

Primary Rating

4.7
Not enough ratings

Secondary Ratings

Novelty
-
Significance
-
Scientific rigor
-
Rate this paper

Recommended

No Data Available
No Data Available