Understanding Routable PCIe Performance for Composable Infrastructures

Explore a comprehensive analysis of routable PCIe performance in composable infrastructures through this 17-minute conference talk from NSDI '24. Dive into the design and implementation of rPCIeBench, a software-hardware co-designed benchmarking framework that systematically characterizes routable PCIe fabric. Learn about flexible data communication primitives, end-to-end PCIe transaction observability, and reconfigurable experiment deployment. Examine the communication characteristics of routable PCIe paths, quantify performance tax, and compare it to local PCIe links. Discover insights into in-fabric traffic orchestration behaviors, including approximate max-min bandwidth partition, fast end-to-end bandwidth synchronization, and interference-free orthogonal data paths. Understand how these characterization insights are encoded as traffic orchestration rules and explore an edge constraints relaxing algorithm for estimating PCIe flow transmission performance over shared fabric. Gain valuable knowledge on optimizing flow schedulers and improving efficiency in composable infrastructures.

NSDI '24 - Understanding Routable PCIe Performance for Composable Infrastructures