On Fri Nov 21 12:31:13 EST 2014, ba...@bitblocks.com wrote:
> This paper is well worth reading:
> http://groups.csail.mit.edu/ana/Publications/PubPDFs/1988Analysis%20TCP%20Processing%20Overhead.pdf
>
> While the traditional BSD implementation uses mbufs that complicate things,
> actual tcp processing can be done quite cheaply.
- ignores tcp checksum -- it alone will take a couple cycles per byte.
- ignores locking
and i believe they've used the BKL to avoid locking any data
structures, otherwise
how could they process ip in 61 instructions? actually there's proof
of this in the
timer instruction count -- 17. that's not enough to acquire a lock.
- asserts that 300 instructions of x86 code -> 400 instructions of risc code,
conservatively
absolutely not if one of them is rep; movb (which they appear to use)
in short, i see signs that this paper is not realistic.
furthermore, this sunny picture assumes an environment where tcp isn't giving
any
benefit. if you're not retransmitting a bit, you're not getting anything out
of tcp.
i think this was the original point.
- erik
