[PATCH 0/21]: powerpc/cell spidernet bugfixes, etc.
James K Lewis
jklewis at us.ibm.com
Thu Oct 12 02:42:51 EST 2006
Please don't be confused by the numbers in the 4M column, they don't
mean anything to the end user. We had a bit better performance (approx.
720 Mbps) at one time but at 100% CPU usage. These new patches lower the
max. to about 700 Mbps but decrease CPU usage down to about 30% which I
think will help all around. I have been testing this driver all night and
have not found a problem yet. Consider this my Acked-by: for all these
Advisory Software Engineer
IBM Linux Technology Center
Arnd Bergmann <arnd at arndb.de>
10/11/2006 11:02 AM
Linas Vepstas <linas at austin.ibm.com>
akpm at osdl.org, jeff at garzik.org, James K Lewis/Austin/IBM at ibmus,
netdev at vger.kernel.org, linux-kernel at vger.kernel.org,
linuxppc-dev at ozlabs.org
Re: [PATCH 0/21]: powerpc/cell spidernet bugfixes, etc.
On Tuesday 10 October 2006 22:49, Linas Vepstas wrote:
> Andrew, please apply/forward upstream.
> The following set of 21 patches (!) are all aimed at the the
> spidernet ethernet device driver. The spidernet is an etherenet
> controller built into the Toshiba southbridge for the PowerPC Cell
> processor. (This is the only device in existance that with this
> ethernet hardware in it).
> These patches re-package/re-order/re-cleanup a previous
> set of patches I've previously mailed. Thus, some have
> been previously Acked-by lines, most do not. Most of
> these patches are tiny, and handle problems that cropped
> up during testing. Sorry about there being so many of them.
> The first set of 12 patches fix a large variety of mostly
> minor bugs.
> The important patches are 13 through 17: these overcome a
> debilitating performance problem on transmit (6 megabits
> per second !!) on transmit of patches 500 bytes or larger.
> After applying these, I am able to get the following:
> pkt sz speed (100K buffs) speed (4M buffs)
> ------ ----------------- ----------------
> 1500 700 Mbits/sec 951 Mbits/sec
> 1000 658 Mbits/sec 770
> 800 600 648
> 500 500 500
> 300 372 372
> 60 70 70
> Above buf size refers to /proc/sys/net/core/wmem_default
Excellent work! I guess this the best tx performance we've
seen so far on this hardware.
Consider this as an Acked-by: for all the patches, I'll save
the effort of replying to each one of them separately.
Jeff, do you plan on merging these fixes for 2.6.19?
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the Linuxppc-dev