Application and Network Performance of Amazon Elastic Compute Cloud Instances

The computing and networking infrastructure in public clouds is shared between multiple users and can create abnormal variations in the performance of applications running in cloud. In this paper we compare the performance of compute-intensive applications on CPUs and GPUs offered by Amazon's cloud and show that not all applications exhibit speedups when executed on the GPU. Even for applications that exhibit speedup on the GPU, the overall application performance may be bottlenecked by network delay. For such cases, the high cost of GPU instances is inefficient since it does not improve application performance. We also show that the network performance of different instances can vary significantly overtime and public clouds may throttle network bandwidth for applications that generate significant network traffic.