Ah, try that on different pieces of h/ware. i.e 32/33, 64/66 and compare
before and after results. You may be suprised by the oddity;->
Actually, why not use patch i supplied?
cheers,
jamal
On Sat, 2004-11-27 at 17:06, Lennert Buytenhek wrote:
> (cc'ing netdev and e1000-devel)
>
>
> On Thu, Nov 25, 2004 at 11:49:45PM -0500, jamal wrote:
>
> > Would be nice to see what happens when you kick the DMA only after
> > filling up the ring.
>
> Just to 'confirm' what you've been seeing, even this tiny little patch:
>
> diff -urN e1000.orig/e1000_main.c e1000/e1000_main.c
> --- e1000.orig/e1000_main.c 2004-11-24 15:35:23.000000000 +0100
> +++ e1000/e1000_main.c 2004-11-27 22:57:08.429782838 +0100
> @@ -1720,7 +1735,9 @@
> wmb();
>
> tx_ring->next_to_use = i;
> + if ((i % 4) == 0) {
> E1000_WRITE_REG(&adapter->hw, TDT, i);
> + }
> }
>
>
> changes my 60B TX rate numbers from:
> 612597pps
> 612447pps
> 612381pps
> 612568pps
> 612579pps
>
> to:
> 614158pps
> 614179pps
> 614186pps
> 614169pps
> 614171pps
> 614150pps
> 614180pps
> 614163pps
> 614179pps
>
> (As before, 32/66, 82541PI, SMP Xeon 2.4, 2.6.9-something.) 'tis not
> much but definitely makes _some_ difference.
>
>
> cheers,
> Lennert
>
|