Re: [PATCH RFC 0/2] SUNRPC: dispatch ready transports round-robin across clients

["Chuck Lever" <cel@kernel.org>]

On Mon, Oct 5, 2026, at 10:46 AM, Benjamin Coddington wrote:
> I know all three of you said no tunables in June.  I'd put this closer to
> exports than to a tunable: it tells the server who is who.  Neil floated
> an export flag for marking clients on the first thread.  I'd do it through
> the nfsd netlink interface so it's per-net and not a module parameter.

I'm not sure how an admin would set such a tunable for a workload that
has to handle more than one of the fairness behaviors.

Scheduling, like getting old, is hard.


> It does raise the bar on the cost question.  If nothing is configured then
> everyone pays for the two-level queue and gets nothing for it, so the
> answer on your "alone" numbers has to be zero, not small.
>
> Would you take something like that?  I'd rather ask before I build it.

The "take/reject" question is answered by looking at patches and test
results, neither of which exist before a prototype is built. The best I
can do a priori is to say "that looks promising or that might suck."

I'll try to think about it more.


-- 
Chuck Lever (Come to NFS bake-a-thon! https://nfsv4bat.org)