Significantly reducing MPI intercommunication latency and power overhead in both embedded and HPC systems

  • Authors:

  • Pavlos M. Mattheakis;Ioannis Papaefstathiou

  • Affiliations:

  • Telecommunication Systems Institute, Technical University of Crete, and University of Crete, Heraklion, Greece;Synelixis Solutions Ltd, Chalkida, Greece

  • Venue:
  • ACM Transactions on Architecture and Code Optimization (TACO) - Special Issue on High-Performance Embedded Architectures and Compilers
  • Year:
  • 2013

Quantified Score

Hi-index 0.00

Visualization

Abstract

Highly parallel systems are becoming mainstream in a wide range of sectors ranging from their traditional stronghold high-performance computing, to data centers and even embedded systems. However, despite the quantum leaps of improvements in cost and performance of individual components over the last decade (e.g., processor speeds, memory/interconnection bandwidth, etc.), system manufacturers are still struggling to deliver low-latency, highly scalable solutions. One of the main reasons is that the intercommunication latency grows significantly with the number of processor nodes. This article presents a novel way to reduce this intercommunication delay by implementing, in custom hardware, certain communication tasks. In particular, the proposed novel device implements the two most widely used procedures of the most popular communication protocol in parallel systems the Message Passing Interface (MPI). Our novel approach has initially been simulated within a pioneering parallel systems simulation framework and then synthesized directly from a high-level description language (i.e., SystemC) using a state-of-the-art synthesis tool. To the best of our knowledge, this is the first article presenting the complete hardware implementation of such a system. The proposed novel approach triggers a speedup from one to four orders of magnitude when compared with conventional software-based solutions and from one to three orders of magnitude when compared with a sophisticated software-based approach. Moreover, the performance of our system is from one to two orders of magnitude higher than the simulated performance of a similar but, relatively simpler hardware architecture; at the same time the power consumption of our device is about two orders of magnitude lower than that of a low-power CPU when executing the exact same intercommunication tasks.