Conclusion

This paper focuses on features of the Trapeze messaging system that support data-intensive cluster OS services, and their use in the GMS network memory system. The paper makes three contributions:

• It describes useful techniques for copy-free handling of page and block transfers in a network storage system, including an incoming payload table (IPT) on the NIC, RPC stub support for zero-copy replies using the IPT, and unified buffering in the network, file, and virtual memory subsystems.
• It shows how to implement useful RPC variants for peer-peer OS services, including delegated RPC and nonblocking RPC using continuations.
• It illustrates use of these techniques in a network memory system that meets aggressive performance goals on a gigabit Myrinet network, using the file system interface to source and sink data at close to network and I/O bus speeds. The GMS/Trapeze prototype features integrated support for high-speed network storage at three levels of the system: (1) the network interface firmware, (2) file/VM and networking subsystems, and (3) the system call interface, with optimizations for the memory-mapped block file access scheme.