Low-latency communication is one area we’ve been exploring within the Office of the CTO as part of our efforts to enable new workloads to run well on the platform, including High Performance Computing (HPC) applications and scale-out Enterprise applications. If you are interested in hearing about this topic at VMworld 2012, consider casting a vote for the two proposals listed below during the public voting period, which ends on June 8th. Vote here (free registration required).
1514 Ultra-low latency on vSphere: How low can we go?
Low-latency applications are becoming more important in the Enterprise datacenter and for emerging workloads like High Performance Computing. As a response to this, the Office of the CTO is exploring several options to deliver the lowest possible latencies on vSphere, starting initially with DirectPath I/O based solutions and looking longer term at building a virtualized RDMA device to deliver low latencies while maintaining platform flexibility. In this presentation we will present performance results using RDMA both in a guest to achieve low application latencies (under two microseconds) as well as within ESXi itself as an experimental demonstration of the value of low-latency services within vSphere for accelerating vMotion operations. We will close the presentation with a description of an advanced development project aimed at creating a virtualized device to provide RDMA and its benefits without requiring the use of Direct Path I/O.
1983 Running latency critical technical applications on vSphere: Performance results
Interest in virtualizing latency-sensitive technical applications is increasing and low-latency requirements in the modern, scale-out datacenter continue to grow. It is therefore important that we assess the ability of ESXi to support such workloads by testing such real-world applications on the vSphere platform. The results of such testing provide guidance to customers and partners regarding best practices for deploying such applications in a virtualized environment and set expectations for the performance that may currently be achieved with such applications.
We have evaluated ESXi performance using a variety of technical benchmarks and applications. The tests have been chosen to be representative of a wide range of technical application types and to also provide helpful information for deploying any type of extremely latency-sensitive application, whether in the realm of technical computing or not.
These results represent joint work between the VMware Office of the CTO, Intel Corporation, and the University of Pisa.