SLO-aware Colocation of Data Center Tasks Based on Instantaneous Processor Requirements

摘要

In a cloud data center, a single physical machine simultaneously executesdozens of highly heterogeneous tasks. Such colocation results in more efficientutilization of machines, but, when tasks' requirements exceed availableresources, some of the tasks might be throttled down or preempted. We analyzeversion 2.1 of the Google cluster trace that shows short-term (1 second) taskCPU usage. Contrary to the assumptions taken by many theoretical studies, wedemonstrate that the empirical distributions do not follow any singledistribution. However, high percentiles of the total processor usage (summedover at least 10 tasks) can be reasonably estimated by the Gaussiandistribution. We use this result for a probabilistic fit test, called theGaussian Percentile Approximation (GPA), for standard bin-packing algorithms.To check whether a new task will fit into a machine, GPA checks whether theresulting distribution's percentile corresponding to the requested servicelevel objective, SLO is still below the machine's capacity. In our simulationexperiments, GPA resulted in colocations exceeding the machines' capacity witha frequency similar to the requested SLO.