Server technologies

Overview

Our team has over two decades of experience in the area of switching, with deep roots in performance analysis. The IBM PowerPRS Packet Routing Switches [1], (aka Prizma) for the Telco market was a key achievement in the 1990s. In 2003, we switched our focus to fabrics used in high-performance computing (HPC) systems. During the subsequent 4 years, in partnership with Corning, Inc (NY, USA) we completed the OSMOSIS (Optical Shared MemOry Supercomputer Interconnect System) optical switch demonstrator. With this project, sponsored by the US Department of Energy and completed in the fall of 2007, we showed that the use of optical switching for HPC is technically feasible, after having cleared many technical hurdles [2]. Our team focused on system design, the design and implementation of the central scheduler for this bufferless crossbar, and performance analysis.

Over the years, we have adapted our performance analysis methods to address the challenges and needs for HPC. We have added the capability to replay MPI (message passing interface) traffic to drive our simulations for HPC interconnect fabrics. We also shifted our focus from highly optimized single stage switch architectures for the Telco industry to multistage interconnect fabrics used in data centers.

We also work on commercial data center interconnects with a focus on Ethernet technology. We have made significant contributions to the IEEE 802.1Qau task group working on congestion control for 10 Gb/s link speeds and up.

In summary, the common theme of our activities is “data motion in data centers”.