Dependence-driven techniques in system design

Burstiness in workloads is often found in multi-tier architectures, storage systems, and communication networks. This feature is extremely important in system design because it can significantly degrade system performance and availability. This dissertation focuses on how to use knowledge of burstiness to develop new techniques and tools for performance prediction, scheduling, and resource allocation under bursty workload conditions.;For multi-tier enterprise systems, burstiness in the service times is catastrophic for performance. Via detailed experimentation, we identify the cause of performance degradation on the persistent bottleneck switch among various servers. This results in an unstable behavior that cannot be captured by existing capacity planning models. In this dissertation, beyond identifying the cause and effects of bottleneck switch in multi-tier systems, we also propose modifications to the classic TPC-W benchmark to emulate bursty arrivals in multi-tier systems.;This dissertation also demonstrates how burstiness can be used to improve system performance. Two dependence-driven scheduling policies, SWAP and ALoC, are developed. These general scheduling policies counteract burstiness in workloads and maintain high availability by delaying selected requests that contribute to burstiness. Extensive experiments show that both SWAP and ALoC achieve good estimates of service times based on the knowledge of burstiness in the service process. as a result, SWAP successfully approximates the shortest job first (SJF) scheduling without requiring a priori information of job service times. ALoC adaptively controls system load by infinitely delaying only a small fraction of the incoming requests.;The knowledge of burstiness can also be used to forecast the length of idle intervals in storage systems. In practice, background activities are scheduled during system idle times. The scheduling of background jobs is crucial in terms of the performance degradation of foreground jobs and the utilization of idle times. In this dissertation, new background scheduling schemes are designed to determine when and for how long idle times can be used for serving background jobs, without violating predefined performance targets of foreground jobs. Extensive trace-driven simulation results illustrate that the proposed schemes are effective and robust in a wide range of system conditions. Furthermore, if there is burstiness within idle times, then maintenance features like disk scrubbing and intra-disk data redundancy can be successfully scheduled as background activities during idle times