题目: TOWARDS PREDICTABLE AND EFFICIENT CLOUD COMPUTING
As virtualization is becoming increasingly ubiquitous in datacenters, there is a growing interest in understanding application performance in multi-tenant systems. Despite that a large body of literature has focused on characterizing virtualization overhead, there still lacks a true understanding of why application performance is often quite poor and unpredictable in virtualized environments. Since virtualization solutions such as hypervisors and containers are in constant development for decades, their overhead alone does not account for the large degree of performance degradations and variations. The gap between virtual and physical systems presents an important obstacle to adopting virtualization in many critical domains and consolidating workloads for datacenter efficiency.
This talk discusses the challenges of delivering predictable performance and attaining high efficiency in multi-tenant clouds and presents our ongoing work to address these issues. First, we present a systematic study on the performance of multi-threaded applications under resource contention and identify the sources of unpredictability. To this end, we develop an online approach to predicting parallel performance under interference. Second, we identify resource discontinuity as one of the inherent gaps between virtual and physical systems and find that the gap can render many resource management policies in the guest OS ineffective. Our research discovers that there exist short-term and long-term priority inversions in the guest OS when virtual machine perceived time becomes discontinuous, which can lead to suboptimal and unpredictable I/O performance. To address this issue, we develop a simple approach based on the concept of CPU ballooning to preserve the static and dynamic priorities in Linux guests.
Dr. Jia Rao is currently an Assistant Professor in the Department of Computer Science at the University of Colorado, Colorado Springs. His research interests include Operating Systems, Virtualization, Cloud computing, Parallel and Distributed Systems, and Machine learning. His research on datacenter efficiency and BigData processing has won one Best Paper award in ICAC, and two Best Paper nominations in HPCA and HPDC. Dr. Jia Rao received his B.S. and M.S. degrees in Computer Science from Wuhan University in 2004 and 2006, respectively, and a Ph.D. degree in Computer Engineering from Wayne State University in 2011.