On Fri, Dec 16, 2005 at 09:40:48AM -0800, Stephen Hemminger ([EMAIL PROTECTED]) wrote: > On Fri, 16 Dec 2005 11:39:52 +0300 > Evgeniy Polyakov <[EMAIL PROTECTED]> wrote: > > > On Thu, Dec 15, 2005 at 12:55:48PM -0800, Stephen Hemminger ([EMAIL > > PROTECTED]) wrote: > > > Given the pretty extensive modifications necessary, do you have any > > > useful benchmark data to show that this is a win. Previous zero > > > copy (like page flipping) has not be merged because it did not > > > prove to be a net gain. > > > > Yes, I have one. > > I use high-performance memory mapped 8139 adapter (original idea > > was to reduce CPU usage on my home gateway when downloading huge files > > over ftp where I use such an Adapter) on 2.4 Ghz Xeon (1+HT): > > 8139 and high performance are an oxymoron. > Any hardware that can't do real DMA is not really high performance.
I know, that is why Adapter is strted from capital letter. This is not about high performance, since current processors easily fill even gigabit link with any packet size, but about cpu usage reduce. The most expensive is ioread according to oprofile, which can not be changed. The second one is memory allocation and then copying to userspace. My approach removes the last one and saves 20% of recv/write CPU usage and 5% in absolute numbers. > If you want to work on a way to do faster receives, buy a better NIC It is not a problemm but this is an expansion way, and I like my old noisy router ;) > then if you really feel talented, make AIO work for network receives. It is already in a some way. The most critical path between network interrupt and protocol processing is splitted. Socket layer is huge, but it is not the most expensive part. Do you think context switches and locking have such a big value in a multi-threaded network server design? > -- > Stephen Hemminger <[EMAIL PROTECTED]> > OSDL http://developer.osdl.org/~shemminger -- Evgeniy Polyakov - To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html
