[openib-general] Re: [PATCH] repost: IPoIB queue size tune patch

[Shirley Ma]

"Michael S. Tsirkin" <mst at mellanox.co.il> wrote on 04/04/2006 12:22:16 PM:

> Quoting r. Shirley Ma <xma at us.ibm.com>:
> > IPoIB is not a real driver. Packets can be dropped from dev xmit queue 
to
> > QPs send queue directly. Why we use tx_ring to induce overhead in
> > send path?
> 
> As I see it, rx_ring and tx_ring are mostly to keep the dma mapping.
> The fact that we have to track it implies I think a level of indirection
> and since the number of these matches the qp size, it seems best to
> allocate them at interface up and not dynamically.
> But: by all means please go ahead and experiment: I was just trying 
> to help with
> these suggestions - I'll guess we'll see what you come up with.

Sure. I will submit the patch for review when I've done all the test.

> In particular, lots of people recommend looking at NAPI as a way
> to reduce iterrupt rate on receive.

I have several experimental patches which do reduce the interrupts/
cpu utilization and increase throughput. These patches are not 
mature enough to submit yet.

In the meantime, we are looking at NAPI. 

Thanks
Shirley Ma
IBM Linux Technology Center
15300 SW Koll Parkway
Beaverton, OR 97006-6063
Phone(Fax): (503) 578-7638
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openfabrics.org/pipermail/general/attachments/20060404/65577b23/attachment.html>