Fast and QoS-Aware Heterogeneous Data Center Scheduling Using Locality Sensitive Hashing

As cloud becomes a cost effective computing platform, improving its utilization becomes a critical issue. Determining an incoming application's sensitivity toward various resources is one of the major challenges to obtain higher utilization. To this end, previous research attempts to characterize an incoming application's sensitivity toward interference on various resources (Source of Interference or SoI, for short) of a cloud system. Due to time constraints, the application's sensitivity is profiled in detail for only a small number of SoI, and the sensitivities for the remaining SoI are approximated by capitalizing on knowledge about some of the applications (i.e. training set) currently running in the system. A key drawback of previous approaches is that they have attempted to minimize the total error of the estimated sensitivities, however, various SoI do not behave the same as each other. For example, a 10% error in the estimate of SoI A may dramatically effect the QoS of an application whereas a 10% error in the estimate of SoI B may have a marginal effect. In this paper, we present a new method for workload characterization and scheduling that considers these important issues. First, we compute an acceptable error for each SoI based on its effect on QoS, and our goal is to characterize an application so as to maximize the number of SoI that satisfy this acceptable error. Then we present a new technique for workload characterization and scheduling based on Locality Sensitive Hashing (LSH). Given a set of n points in a d-dimensional Euclidean space, LSH is a hashing technique such that points nearby are hashed to the same "bucket" and points that are far apart are hashed to different buckets. This data structure allows approximate nearest neighbor queries to be executed with nearly asymptotically optimal running time. This allows us to perform workload profiling quickly with high accuracy and scheduling in heterogeneous data centers with high quality of service (QoS) and utilization.