Architecting Data Centers for High Efficiency and Low Latency

data center;efficiency;tail latency;Computer Science;Engineering;Computer Science & Engineering

Zhang, YunqiWenisch, Thomas F ;

University of Michigan

关键词: data center;    efficiency;    tail latency;    Computer Science;    Engineering;    Computer Science & Engineering;   

Others  :  https://deepblue.lib.umich.edu/bitstream/handle/2027.42/144144/yunqi_1.pdf?sequence=1&isAllowed=y

瑞士|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

Modern data centers, housing remarkably powerful computational capacity, are built in massive scales and consume a huge amount of energy. The energy consumption of data centers has mushroomed from virtually nothing to about three percent of the global electricity supply in the last decade, and will continuously grow. Unfortunately, a significant fraction of this energy consumption is wasted due to the inefficiency of current data center architectures, and one of the key reasons behind this inefficiency is the stringent response latency requirements of the user-facing services hosted in these data centers such as web search and social networks. To deliver such low response latency, data center operators often have to overprovision resources to handle high peaks in user load and unexpected load spikes, resulting in low efficiency.This dissertation investigates data center architecture designs that reconcile high system efficiency and low response latency. To increase the efficiency, we propose techniques that understand both microarchitectural-level resource sharing and system-level resource usage dynamics to enable highly efficient co-locations of latency-critical services and low-priority batch workloads. We investigate the resource sharing on real-system simultaneous multithreading (SMT) processors to enable SMT co-locations by precisely predicting the performance interference. We then leverage historical resource usage patterns to further optimize the task scheduling algorithm and data placement policy to improve the efficiency of workload co-locations. Moreover, we introduce methodologies to better manage the response latency by automatically attributing the source of tail latency to low-level architectural and system configurations in both offline load testing environment and online production environment. We design and develop a response latency evaluation framework at microsecond-level precision for data center applications, with which we construct statistical inference procedures to attribute the source of tail latency. Finally, we present an approach that proactively enacts carefully designed causal inference micro-experiments to diagnose the root causes of response latency anomalies, and automatically correct them to reduce the response latency.