Converging Interconnect Requirements for HPC and Warehouse Scale Computing

7.00-7.30: welcome reception

7.30-8.30: "Converging Interconnect Requirements for HPC and Warehouse Scale Computing"

John Shalf

-- Department Head for Computer Science for the Computing Research Division @ Lawrence Berkeley Lab

-- Chief Technology Officer @ NERSC

This talk will discuss the emerging trends towards convergence of cloud and HPC interconnect requirements due to both market forces and increasingly aligned performance requirements.

One point of divergence between HPC and Cloud computing is the interconnect technology. Energy efficiency concerns and the cost effectiveness of using commodity technology has strongly aligned both HPC and cloud computing in terms of processor and cooling technology, but the interconnects have remained a big differentiating factor. HPC has typically required custom high performance interconnects for tightly-coupled inward-facing traffic. Cloud computing has tended to focus on standards-based Ethernet with TCP/IP for outward-facing wide area networking. However, with the rise of large scale data analytics applications, warehouse scale computing systems have observed that large fractions of their communications traffic is inward-facing, and that the performance of the inward-facing datacenter network has become a primary limiting factor for performance.

Furthermore, recent acquisition of high performance interconnect technologies by Intel for integration directly onto the processor has fundamentally changed the ecosystem for commodity server chips.

The technologies are converging, but will the ends meet? The talk will examine trends in both requirements for datacenter workloads as well as emerging technology trends and what it means for the future of HPC and cloud environments.

8.30-9.00: socializing