Conclusion

   This paper summarizes recent and current research in the Trapeze project at Duke University. The goal of our work has been to push the performance bounds for network storage access using Myrinet networks, by exploring new techniques and optimizations in the network interface and messaging system, and in the operating system kernel components for file systems and virtual memory. Key elements of our approach include:

• An adaptive message pipelining strategy implemented in custom firmware for Myrinet NICs. The Trapeze firmware combines low-latency transfers of I/O blocks with high bandwidth under load.
• A lightweight kernel-kernel RPC layer optimized for network I/O traffic. NetRPC allows zero-copy sends and receives directly from the I/O cache, and supports nonblocking RPCs with interrupt-time reply handling for high-bandwidth prefetching.
• A scalable storage system incorporating network memory and parallel disks on a collection of PC-based I/O servers. The system achieves high bandwidth and capacity by using many I/O nodes in tandem; it can scale incrementally with demand by adding more I/O nodes to the network.

Slice uses Trapeze and NetRPC to access network storage at speeds close to the limit of the network and client I/O bus. We expect that this level of performance can also be achieved with new network standards that are rapidly gaining commodity status, such as Gigabit Ethernet.