Thu, 09 Feb 2006
Van Jacobson's network channels
In DaveM's blog there is an interesting followup to the Van Jacobson's talk (slides in PDF) at linux.conf.au.
Van Jacobson suggests that the kernel networking stack should be reworked as channels (one-way lock-free queues) of packets, and the parsing and handling of the network packets is to be done as near to the end of the "food supply chain" as possible (i.e., in the user-space apps, if possible). He also gives the numbers which show the better scalability of this approach. The scalability is important especially in SMP, NUMA and multi-core systems, which are becoming more and more common these days.
While this approach is definitely interesting, Van Jacobon leaves out an important problem - how the security can be accomplished? When any app is allowed to send arbitrary packets (because it does user-space TCP), how it can be kept from interfering with other apps, disrupting other TCP connections, and so on? DaveM's suggestion is to make "channel-based" TCP in kernel, with a tiny packet classifier, which allows mapping of the packets in the device's input channel to the channel of the particular socket. The TCP handling would then be done in the context of that particular process (yet in kernel space).
Van Jacobson's measurements suggest this way the TCP processing on a SMP box can be 6 times faster (and essentially lock-free) than in the current kernel (while he also acknowledges that Linux net stack already is the fastest and most complete networking stack of any OS). There is also a LWN followup in last week's LWN "Kernel" section.