Van Jacobson Suggeston to improve the linux tcp stack Linux has the fastest networking stack, he is not criticising it, he wants to suggest a way to change something that is in linux for historic reasons predating linux the networking stack goes inside the kernel because Multics had it inside for ARPA Net. The multics stack begat many other stacks, tcp stack went inside the kernel because of this This became the standard model to build stacks, some stuff inside app, some stuff inside kernel the way we have always done it is not necessarily the right way networking based on two loops (user to kernel and kernel to other end point) is going to be less stable than a single loop mechanism. kernel based implementations often have to make an extra copy of data that heads to user space. Copies are expensive. Dave Clark at MIT came up with the principle for making the network and other stuff work as it scales up, be as lazy as you can be, push work downstream to the nodes and dont make the central components do all the work the testing result she has are damn impressive grab the slides to see a lot of the details with what he has been doing starting to be severely limited by the DDR 333 memory speed with cpus sitting idle rather than too many interrupts and such in cpu slowing things down His conculsion sugests there is a way to get linear improvement on multi cpu systems