Inferring the Topology and Traffic Load of Parallel Programs Running in a Virtual Machine Environment

We are developing a distributed computing environment based on virtual machines featuring application monitoring, network monitoring, and an adaptive virtual network. In this paper, we describe our initial results in monitoring the communication traffic of parallel applications, and inferring its spatial communication properties. The ultimate goal is to be able to exploit such knowledge to maximize the parallel efficiency of the running parallel application by using VM migration, virtual overlay network configuration and network reservation techniques, which are a part of the distributed computing environment. Specifically, we demonstrate that: (1) we can monitor the parallel application network traffic in our layer 2 virtual network system with very low overhead, (2) we can aggregate the monitoring information captured on each host machine to form a global picture of the parallel application's traffic load matrix, and (3) we can infer from the traffic load matrix the application topology. In earlier work, we have demonstrated that we can capture the time dynamics of the applications. We begin here by considering offline traffic monitoring and inference as a proof of concept, testing it with a variety of synthetic and actual workloads. Next, we describe the design and implementation of our online system, the Virtual Topology and Traffic Inference Framework (VTTIF), and evaluate it using a NAS benchmark.