Enabling Flexible Collective Communication Offload with Triggered Operations
Authors:
- Keith D. UnderwoodKUKeith D. Underwood
- Jerrie CoffmanJCJerrie Coffman
- Roy LarsenRLRoy Larsen
- K. Scott HemmertKHK. Scott Hemmert
- Brian W. BarrettBBBrian W. Barrett
- Ron BrightwellRBRon Brightwell
- Michael LevenhagenMLMichael Levenhagen
- 1 August 2011
- conference paper
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- No. 15504794,p. 35-42
Abstract
Low latency collective communications are key to application scalability. As systems grow larger, minimizing collective communication time becomes increasingly challenging. Offload is an effective technique for accelerating collective operations, however, algorithms for collective communication constantly evolve such that flexible implementations are critical. This paper presents triggered operations -- a semantic building block that allows the key components of collective communications to be offloaded while allowing the host side software to define the algorithm. Simulations are used to demonstrate the performance improvements achievable through the offload of MPI_Allreduce using these building blocks.Keywords
This publication has 13 references indexed in Scilit:
- An application based MPI message throughput benchmark2009
- Characterizing application sensitivity to OS interference using kernel-level noise injection2008
- Simulating Red Storm: Challenges and Successes in Building a System Simulation2007
- Optimised Global Reduction on QsNet^ⅠⅠ2006
- NIC-based offload of dynamic user-defined modules for Myrinet clusters2005
- Efficient Implementation of Allreduce on BlueGene/L Collective Network2005
- Optimization of Collective Reduction Operations2004
- Where to provide support for efficient multicasting in irregular networks: network interface or switch?2002
- Fast NIC-based barrier over Myrinet/GM2002
- Optimized routing in the Cray T3D1994