MadSystems Seminar -- Yuhong Zhong (Columbia)

Title: The Next Frontier in Efficiency: Pooling PCIe Devices with CXL

Abstract: Compute Express Link (CXL) has already been deployed in production as a key technology for scaling memory capacity. As the next step in this evolution, CXL memory pools allow hosts to dynamically allocate memory from the pool based on demand, improving memory utilization in datacenters. In this talk, I will present our recent work showing that, once CXL memory pools are in place, they also enable pooling PCIe devices. In addition to CPU and memory, PCIe devices such as NICs and SSDs are often significantly underutilized in datacenters. Pooling PCIe devices, where multiple hosts share a common set of devices, can improve utilization and reduce total cost of ownership. Today, this is primarily achieved using PCIe switches, which are expensive and inflexible. We design Oasis, a system that enables PCIe device pooling in software by leveraging CXL memory pools. Once deployed, CXL pools can serve as an efficient data path between hosts and PCIe devices. Oasis provides both a control plane and a datapath over CXL, mapping and routing PCIe device traffic across host boundaries. As a proof of concept, we implement an Oasis network engine to enable NIC pooling. Our evaluation shows that Oasis improves NIC utilization by 2× while supporting fast NIC failover with only a 38 ms interruption.

Bio: Yuhong Zhong is a fourth-year PhD candidate in Computer Science at Columbia University, advised by Asaf Cidon. His recent research is driven by an ambition to unlock the full potential of CXL to improve CPU, memory, and PCIe device efficiency in datacenters, with a focus on memory tiering and memory and PCIe device pooling. His work was recognized with the Best Paper Award at OSDI '22. Prior to his PhD, he worked as a software engineer at VMware on the vSAN team.

Link: https://uwmadison.zoom.us/j/99445268808?pwd=feYjJEbqyiT8Q2yf3IbgYPyT5xPYVB.1