Robert Watson's Slides from MeetBSD 2008

MeetBSD 2008 took place in Mountain View, California, at the Googleplex, and included a variety of BSD-related talks, a FreeBSD developer summit, and of course the 15th anniversary party for FreeBSD. On this page, you can find my slides from th conference and developer summit.

200810-meetbsd-networking.pdf - FreeBSD Network Stack Optimizations for Modern Hardware (MeetBSD 2008 Full Conference)

The arrival of high CPU core density, with commodity quad-core notebooks and 32-core servers, combined with 10gbps networking have transformed network design principles for operating systems. This talk will describe changes in the FreeBSD 6.x, 7.x, and forthcoming 8.x network stacks required to exploit multiple cores and serve 10gbps networks. The goal of the session will be to introduce the audience to general strategies used to improve performance, their rationales, and their impact on applications and users:

Introduction to the SMPng Project and the follow-on Netperf Project
Workloads and performance measurement
Efficient primitives to support modern network stacks
Multi-core and cache-aware network memory allocator
Fine-grained network stack locking
Load-balancing and contention-avoidance across multiple CPUs
CPU affinity for network stack data structures
TCP performance enhancements including TSO, LRO, and TOE
Zero-copy Berkely Packet Filter (BPF) buffers
Direct network stack dispatch from interrupt handlers
Multiple input and output queues

20081116-meetbsd-foundation.pdf - FreeBSD Foundation Status Update (MeetBSD 2008 Full Conference)

A brief update on the status of the FreeBSD Foundation, including recent activities, fund-raising, and developer recognition.

20081117-devsummit-networking.pdf - Update on the SMP Network Stack (MeetBSD 2008 FreeBSD Developer Summit)

A brief update on on-going activities and areas of optimization for the multi-processor network stack project. We review recent accomplishments, works-in-progress, some thing that need to happen urgently, and thoughts on changing the socket API to improve application and kernel scalability.