An execution context optimization framework for disk energy

Power, energy, and thermal concerns have had explosive growth in research over the past two decades. In servers, desktops, and mobile systems, the hard disk is among the top resources in power and energy consumption. Common techniques for reducing disk energy consumption have included caching, adaptive low power modes, batch scheduling, and data migration. Many previous software optimizations for single disk systems have assumed and experimented in uniprogramming environments. However, modern systems are typically multiprogramming, and the optimizations do not extend well from the uniprogramming model. Programs should be aware of concurrently running programs to enable cooperation and coordinate disk accesses from multiple programs. The set of concurrently running programs is referred to as an execution context. Execution context optimizations were introduced to target multiprogramming environments. My research introduces an optimization framework to provide execution context information and reduce disk energy consumption by effectively managing disk accesses. Optimizing over all possible execution contexts is counter-productive because many contexts do not occur in practice. For an extreme example, users rarely, if at all, run more than twenty programs concurrently. Optimizations may be profitably targeted at the most common execution contexts for a given workload. A study was conducted of real workloads by collecting user activity traces and characterizing the execution contexts. Out of hundreds of contexts and over 50 unique programs, the study confirmed the intuition that users generally run only a small set of programs at a time. Execution context optimizations were implemented on eight streaming and interactive applications. The optimizations were compared to previous best optimizations and evaluated on a laptop disk which is already designed for energy efficiency. The disk energy was measured while running synthetic traces of ten execution contexts. The results show up to 63% energy savings while incurring less than 1% performance delay. When compared to unoptimized versions, energy savings was up to 77%. If the optimizations were applied to comparable applications in the user study, an estimated 9% disk energy could have been saved. Execution context optimizations show significant promise for saving disk energy.Ph.D.Includes bibliographical references (p. 89-95)