Abstract:

With the development of cloud computing, cloud file systems, such as GFS and HDFS, play an increasingly significant role within the backend of cloud infrastructures. Various benchmarks for evaluating the performance of cloud file systems have been proposed. However, most of these benchmarks focus on the traditional performance metrics, like input/output operations per seconds (IOPS) and throughput, while ignoring the measurement of the performance isolation in multi-tenant environments. Because the I/O workloads in clouds are heterogeneous and dynamic, it is a challenge to measure performance isolation of cloud file systems. We propose a novel model for measuring the isolation, and implement this model in a benchmark suite, called Porcupine. The Porcupine schedules request streams following a workload-specific arrival pattern. These request sequences conform to the characteristics  of workloads in cloud file systems. Porcupine can be used to deploy benchmark test on cloud file systems conveniently. Experimental results demonstrate that isolation measurement model is effective.

Key words: performance isolation, cloud file systems, benchmark, multi-tenant environment