Speaker: Jeff Chase, Duke University
Date: January 21, 1999
Cheating the I/O Bottleneck: Network Storage with Trapeze/Myrinet
The goal of the Trapeze project is to realize the potential for high-speed data transfer over Myrinet networks, and to harness this potential for cluster file systems, network memory systems, and other services that cooperatively share data across a cluster. Our broad goal is to use the power of gigabit networks to "cheat" the disk I/O bottleneck for data-intensive computing, by taking a network-centric view of I/O in which the network is the primary access path to I/O resources. This work brings together aspects of network interface design, file and virtual memory systems, I/O prefetching, and operating system kernel structure for high-speed communication.
The talk will describe our approach to adaptive message pipelining in the Trapeze network interface, which balances latency and bandwidth for page/block transfers characteristic of network storage systems. I will also present key features of our network memory prototype, based on the University of Washington Global Memory Service (GMS), and of its Trapeze-based kernel-kernel RPC layer. Our latest GMS/Trapeze prototype can service 8KB page faults from network memory in 165 microseconds, with peak sequential file access bandwidths of 100 MB/s.