On Tue, Jan 20, 2009 at 12:01:13PM +0000, Ben Mansell (ben@zeus.com) wrote:

I've also tested on the same three patches (against 2.6.27.2 here), and 
the patches appear to work just fine. I'm running a similar proxy 
benchmark test to Willy, on a machine with 4 gigabit NICs (2xtg3, 
2xforcedeth). splice is working OK now, although I get identical results 
when using splice() or read()/write(): 2.4 Gbps at 100% CPU (2% user, 
98% system).

With small MTU or when driver does not support fragmented allocation
(iirc at least forcedeth does not) skb will contain all the data in the
linear part and thus will be copied in the kernel. read()/write() does
effectively the same, but in userspace.
This should only affect splice usage which involves socket->pipe data
transfer.

I may be hitting a h/w limitation which prevents any higher throughput, 
but I'm a little surprised that splice() didn't use less CPU time. 
Anyway, the splice code is working which is the important part!

Does splice without patches (but with performance improvement for
non-blocking splice) has the same performance? It does not copy data,
but you may hit the data corruption? If performance is the same, this
maybe indeed HW limitation.

-- 
	Evgeniy Polyakov