Mitigating the cost, performance, and power overheads induced by load variations in multicore cloud servers
Abstract
Load variations whether in space or time pose a significant challenge to system designers. These load variations may induce inefficiencies such as load imbalance and over-provisioning, resulting in performance/power/cost overheads. The goal of my research is to mitigate such variation-induced overheads in multicore cloud servers. First, I focus on power/performance overheads in on-chip networks of a multicore chip. We design an on-chip network that is robust in both performance and energy across applications for time- and space-varying loads. Existing flow control mechanisms that perform well at high (low) loads suffer power and/or energy overheads at low (high) loads. In contrast, our design dynamically adapts flow control to achieve power and performance of the better-suited flow-control mechanism at all loads. Second, I target cost overheads resulting from time-varying loads for applications hosted in an Infrastructure-as-a-Service (IaaS) cloud. While IaaS clouds may enable significant cost-savings by allowing elastic provisioning, the uncertainty of time-varying loads impose additional cost to maintain quality of service. I demonstrate that, with some knowledge of the statistical properties of time-varying load, one can maximize cost-savings while satisfying response-time targets. Finally, I propose to mitigate the impact of data popularity variations in cloud servers. Sharding is a common technique to partition data among scale-out servers. Unfortunately, skewed popularity of data-elements can cause significant load imbalance among shard servers, leading to response time degradation. I design an augmented variant of a well-known memory-caching system to identify and replicate popular read-mostly data elements, thus achieving better load balance and higher performance.
Degree
Ph.D.
Advisors
Thottethodi, Purdue University.
Subject Area
Computer Engineering
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.