Analysis and Modeling of Social In uence in High Performance Computing Workloads

摘要

High Performance Computing (HPC) is becoming a common tool in many researchareas. Social influence (e.g., project collaboration) among increasing users of HPCsystems creates bursty behavior in underlying workloads. This bursty behavior isincreasingly common with the advent of grid computing and cloud computing. Miningthe user bursty behavior is important for HPC workloads prediction and scheduling,which has direct impact on overall HPC computing performance.A representative work in this area is the Mixed User Group Model (MUGM),which clusters users according to the resource demand features of their submissions,such as duration time and parallelism. However, MUGM has some difficulties whenimplemented in real-world system. First, representing user behaviors by the featuresof their resource demand is usually difficult. Second, these features are not alwaysavailable. Third, measuring the similarities among users is not a well-defined problem.In this work, we propose a Social Influence Model (SIM) to identify, analyze,and quantify the level of social influence across HPC users. The advantage of theSIM model is that it finds HPC communities by analyzing user job submission time, thereby avoiding the difficulties of MUGM. An offline algorithm and a fast-converging,computationally-efficient online learning algorithm for identifying social groups areproposed. Both offline and online algorithms are applied on several HPC and gridworkloads, including Grid 5000, EGEE 2005 and 2007, and KAUST SupercomputingLab (KSL) BGP data. From the experimental results, we show the existence of a socialgraph, which is characterized by a pattern of dominant users and followers. In orderto evaluate the effectiveness of identified user groups, we show the pattern discoveredby the offline algorithm follows a power-law distribution, which is consistent withthose observed in mainstream social networks. We finally conclude the thesis anddiscuss future directions of our work.