28 Jun
2022
28 Jun
'22
11:32 p.m.
I’ll argue there is quite a lot known about where to put network functionality, much
of it from HPC. If you want minimum latency and minimum variance of latency, both
of which are important to big applications, you make the network reliable and move
functionality into the host adapters and the switches. The code path at each end
of a very good MPI implementation will be under 200 machine instructions, all in
user mode. There is no time to do retry or variable code paths.
Doesn’t work on WANs of course, or at consumer price points.
(I think there is still a lot to do, because the best networks still hover around
800 nanoseconds calling SEND to returning from RECV, and I think it could be 100).
-L
On 2022, Jun 28, at 11:50 AM, Noel Chiappa
<jnc(a)mercury.lcs.mit.edu> wrote:
From: Rob Pike
having the switch do some of the call validation
and even maybe
authentication (I'm not sure...) sounds like it takes load off the host.
I don't have enough information to express a judgement in this particular
case, but I can say a few things about how one would go about analyzing
questions of 'where should I put function [X]; in the host, or in the
'network' (which almost inevitably means 'in the switches')'.