Related Work

In this section we briefly mention representative work in areas where similar approaches are used. Long term schedulers in batch systems have used page fault frequency (PFF) as the objective function for decisions to increase or decrease the multiprogramming level. Our approach is complementary to Morph [17]. Instead of optimizing the performance by rewriting the binary code, we change the behavior of the system software without rewriting its binary code. We argue that it possible for both approaches to applied simultaneously because Morph-like optimizations optimize the code for a specific hardware or end-user pattern, whereas our approach addresses other performance issues, such as thrashing, which should be handled by specific policies (e.g., limiting the level of concurrency or changing the scheduling policy) rather than binary code optimization. Douceur and Bolosky [6] used similar techniques to regulate low-importance processes. Our approach strives to maintain maximal performance by adapting the system behavior using both productivity and overhead metrics. This is in contrast to their approach where only productivity (progress) metrics are used to regulate low-importance processes such that they do not affect the execution of other processes. SEDA [15,14] presents a staged architecture for Internet servers. Our approach is another point in the design space of building systems that change their behavior adaptively to improve performance. SEDA offers greater control within each stage of a server; however, it requires a complete rewrite of the software. In contrast, our approach gives less control and requires far fewer modifications to the software. The MAGNET [7] tool tracks OS events and exports information on them to user level. It has been used to identify Linux kernel problems (e.g., Ethernet driver, scheduler anomalies, overheads) and it has been used to analyze and tune applications, including creation of a reflective application. In the Atlas [16] project, empirical techniques are used to tune the performance of some BLAS and LAPACK routines to provide portable performance. Bershad et al. [3] used feedback from special hardware to dynamically avoid conflict misses in large direct-mapped caches by reassigning the conflicting virtual memory pages. In the AppLeS [2] project, an application-level scheduler is used to adaptively and dynamically schedule individual applications on distributed, heterogeneous systems.