Re: Remote fork() and Parallel programming

Larry McVoy (lm@bitmover.com)
Tue, 16 Jun 1998 13:22:23 -0700


: > Are you sure about the 30,000ns figure? There are several solutions
: > that claim 2-3 microsecond latency using some custom hardware and
: > userspace libraries. These include
:
: He's not refering to banging the hardware from userspace (well, I wasn't
: and he was replying to me)..

Yeah, you're right, I was thinking commodity networking cards. if you
do TCP on 'em, the best numbers I've seen are about 145usecs round trip,
or 77 usecs for one way.

: Though I can't imagine that it needs to be *that* high.. With a special
: (non-ip, non checksummed) ethernet protocall for san's and a highly
: optimized networking path much lower should be possible..

I think you could lose maybe 50 usecs of that 77 due to the protocol
stack, but I think that is pretty damn liberal. The stack isn't really
the issue, it's the interrupt handler and the device driver that gets you.
At least that has been true on things like SunOS and IRIX. The stack is
actually not so bad.

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu