Operating system structure is important; it leads to understandable, maintainable, 'pluggable' code. But despite our best efforts, some system elements have been difficult to structure. We propose a new analysis of this problem, and a new technology that can structure these elements. Aspect-oriented programming (AOP) (G. Kiczales et al., 1997) uses linguistic mechanisms to support the separation of crosscutting elements, or aspects of the system, from primary functionality. We have developed a proof-of-concept AOP implementation of prefetching in FreeBSD (www.cs.ubc.ca/labs/spl/aspects/aspectc.html). In our implementation, we have been able to modularize prefetching.

The benefits of Markov-based predictive prefetching have been largely overshadowed by the overhead required to produce high-quality predictions. While both theoretical and simulation results for prediction algorithms appear promising, substantial limitations exist in practice. This outcome can be partially attributed to the fact that practical implementations ultimately make compromises in order to reduce overhead. These compromises limit the level of algorithm complexity, the variety of access patterns and the granularity of trace data that the implementation supports. This paper describes the design and implementation of GMS-3P (Global Memory System with Parallel Predictive Prefetching), an operating system kernel extension that offloads prediction overhead to idle network nodes. GMS-3P builds on the GMS global memory system, which pages to and from remote workstation memory. In GMS-3P, the target node sends an online trace of an application's page faults to an idle node that is running a Markov-based prediction algorithm. The prediction node then uses GMS to prefetch pages to the target node from the memory of other workstations in the network. Our preliminary results show that predictive prefetching can reduce the remote-memory page fault time by 60% or more and that, by offloading prediction overhead to an idle node, GMS-3P can reduce this improved latency by between 24% and 44%, depending on the Markov model order.

Advances in network technology continue to improve the communication performance of workstation and PC clusters, making high-performance workstation-cluster computing increasingly viable. These hardware advances, however, are taxing traditional host-software network protocols to breaking point. A modern gigabit network can swamp a host's IO bus and processor, limiting communication performance and slowing computation unacceptably. Fortunately, host-programmable network processors used by these networks present a potential solution. Offloading selected host processing to these embedded network processors lowers host overhead and improves latency. This paper examines the use of embedded network processors to improve the performance of workstation-cluster global memory management. We have implemented a revised version of the GMS global memory system that eliminates host overhead by as much as 29% on active nodes and improves page fault latency by as much as 39%.