Understanding Routable PCIe Performance for Composable Infrastructures
Talkat 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI '24) · 2024-04-16..04-18
summary
rPCIeBench is a software–hardware co-designed benchmark that implements four DMA primitives to measure routable PCIe paths using a Gig Fabric X RS4024 switch and FPGA accelerators. It finds a 49.6 ns one‑way latency overhead, bandwidth drops of 30.4% (host→device) and 6.9% (device→host), credit-based run‑robin sharing that yields ~6 GB/s per flow at saturation, and fast bandwidth sync. An age constraints relaxing estimator predicts flow rates with 2.9% average error, validating rPCIeBench.
AI-written summary · about