[openib-general] Re: [PATCH] TX/RX_RING_SIZE as loadable parameters

[Roland Dreier]

    Thomas> I thought UD has only a single-packet window in the qp
    Thomas> context.

Nope.  You're probably thinking about the fact that RD only allows you
to have a single message in flight at a time to a given destination
(but no one has ever really implemented the RD transport anyway).  UD
is just unreliable single packet messages, so once a send request has
been processed, there's no way for the HCA to know when it reaches the
destination, so it's not even theoretically possible to limit UD to a
single message in flight.

    Thomas> Well, it's only important if the code that's there works
    Thomas> well. I thought Shirley said it doesn't. Has anyone
    Thomas> instrumented it for overruns and drops, and watched it
    Thomas> under load? That would tell us what to tweak.

I did a little bit of testing a while ago, and I had a hard time ever
getting the TX ring to fill up with the current default size.

    Thomas> Constants are pretty much never correct in networking
    Thomas> code. And module parameters are darn close to being
    Thomas> constants.

But if we provide a million little knobs, no one can ever tune
things.  For example what's the advantage of having a small TX ring?
Maybe we should just make it big by default.

 - R.