Hello Michael,
"Michael S. Tsirkin" <[EMAIL PROTECTED]> wrote on 04/04/2006 10:22:08 AM:
> Quoting r. Shirley Ma <[EMAIL PROTECTED]>:
> > Subject: Re: [PATCH] repost: IPoIB queue size tune patch
> >
> >
> > Hello Roland,
> >
> > Thanks.
> >
> > Roland Dreier <[EMAIL PROTECTED]> wrote on 04/03/2006 10:50:37 PM:
> >
> > > Sorry I hadn't gotten a chance to read this over until now...
> > >
> > > > - IPOIB_RX_RING_SIZE = 128,
> > > > - IPOIB_TX_RING_SIZE = 64,
> > > > + IPOIB_SENDQ_SIZE = 64,
> > > > + IPOIB_RECVQ_SIZE = 128,
> > >
> > > Can you explain again why it's a good idea to rename these? Is the
> > > name "sendq_size" really clearer than "tx_ring_size," especially in
> > > the context of a network driver?
> >
> > tx_ring and rx_ring are not good names for IPoIB. tx_ring in IPoIBis a place
> > to save pointers and rx_ring a place to allocate skb buffs.
>
> Actually names make sense to me - they are cyclic buffers, hence ring.
>
> > I am working on
> > a patch to remove tx_ring and replace rx_ring with a list, which would
> > reduce some tx_ring overhead.
>
> What overhead? Please don't, lists are bad for cache, require locks ...
> Circular buffers are much better, I don't have the time now but I think we can
> even make event handler completely lockless with them.
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?
Also, using a list doesn't always mean to induce a lock. I will test the patch
performance before I submit for review.
We really want this patch to be in RC2 release. I will change the name back
if this would prevent the patch from being into upper stream.
Thanks
Shirley Ma
IBM Linux Technology Center
15300 SW Koll Parkway
Beaverton, OR 97006-6063
Phone(Fax): (503) 578-7638
_______________________________________________ openib-general mailing list [email protected] http://openib.org/mailman/listinfo/openib-general
To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
