MPI on U-Net

Class Project CS268

Abstract:

A simple protocol is designed and implemented into the ADI layer of the MPICH message passing library to ensure reliability when running on top of unreliable network interfaces such as U-Net. Our protocol for short messages uses a reliable message stream and lazy acknowledgment to realize low latencies and a small protocol overhead. An implementation for two Pentium Pro 200 nodes connected with 100Mb/s Tulip Fast Ethernet shows that latencies of about 32 s can be achieved at the ADI layer.

• Introduction
• Design Decisions
  • Operating Mode
  • Network
• Startup Phase
• Reliability Without Context Switch
• Protocol
  • Description of the Protocol
  • Short Versus Long Messages
  • Sender Versus Receiver Initiated Retransmission
  • Comparison With TCP
• Test Implementation
  • Details of the Implementation
  • Performance Measurements
• Related Work
• Future Work
• Acknowledgments
• References