ron minnich wrote on Sun, Dec 18, 2022 at 08:50:18AM -0800:
> it's fine. tbh, I doubt the fact that you were fetching 31 vs 32 pages
> mattered as much as the fact that you weren't fetching *4k at a time* :-)
Yes, I think we can just blanket this as +4k and it wouldn't change
much; I've been using 1MB+4k for rdma in previous tests...
We still aren't doing things 4k at a time with this though, I'd suggest
rounding down the rsize > msize check in p9_client_read_once():
if (!rsize || rsize > clnt->msize - P9_IOHDRSZ)
rsize = clnt->msize - P9_IOHDRSZ;
to something that's better aligned; for some reason I thought we had
that already. . . but thinking again the sizes are probably driven by
the cache and will be 4k multiples already?
> > -#define DEFAULT_MSIZE (128 * 1024)
> > +/* DEFAULT MSIZE = 32 pages worth of payload + P9_HDRSZ +
> > + * room for write (16 extra) or read (11 extra) operands.
> > + */
> > +
> > +#define DEFAULT_MSIZE ((128 * 1024) + P9_HDRSZ + 16)
There's P9_IOHDRSZ for that ;)
But I guess with the comment it doesn't matter much either way.
--
Dominique
