On PCIe Performance

Add to your list(s) Download to your calendar using vCal

• Rolf Neugebauer (Docker, Inc)
• Thursday 17 March 2016, 15:00-16:00
• FW26, Computer Laboratory, William Gates Building.

If you have a question about this talk, please contact Eiko Yoneki.

Providing network connectivity at 40Gb/s and above is challenging. The PCIe interface used by most 40G NICs is being pushed to the limit for small packets and the DMA engines have to sustain fairly high transaction rates. Host features such as a integrated (uncore) memory and PCIe controllers and tighter integration of the PCIe root complex with the cache subsystem (e.g., Intel’s DDIO ) lead to non-uniform access to host memory for PCIe devices. IO-MMUs and their TLBs are also interposed into the data path, potentially affecting performance.

In this talk, we present a methodology and a set of PCIe micro-benchmarks which are designed to evaluate the impact of these technologies. The micro-benchmarks are currently implemented on Netronome’s NFP -3200, NFP -4000, and NFP -6000 series of network cards and we present detailed data on the impact and the evolution of above technologies over several generations of Intel processors and chipsets. The methodology is more generally applicable and can be used to compare the performance of different DMA engines as well as evaluating different processor architectures.

Bio: Rolf is a Member of Technical Staff at Docker Inc. Previously he was a Distinguished Engineer at Netronome where he led the driver team, responsible for Linux and Windows network device drivers as well as the corresponding firmware for Netronome’s highly programmable 10-100Gb/s NICs. He also worked on a transparent offload/acceleration prototype for Open vSwitch, developed a domain specific language used widely internally during silicon and system software development, and implemented the layer three (IP) control plane for a Next Generation Firewall product. Previously he was a Researcher at Intel Research were he was one the original authors of the Xen hypervisor and also led an internal project on IO virtualisation. Rolf holds a PhD from the University of Glasgow. For his dissertation he applied theories of congestion pricing to resource management in operating systems.

This talk is part of the Computer Laboratory Systems Research Group Seminar series.

This talk is included in these lists:

• All Talks (aka the CURE list)
• CL's SRG seminar
• Cambridge Centre for Data-Driven Discovery (C2D3)
• Cambridge talks
• Chris Davis' list
• Computer Laboratory Systems Research Group Seminar
• Department of Computer Science and Technology talks and seminars
• FW26, Computer Laboratory, William Gates Building
• Interested Talks
• School of Technology
• Trust & Technology Initiative - interesting events
• bld31
• ndk22's list
• ob366-ai4er
• rp587

Note that ex-directory lists are not shown.