Camm Maguire wrote: > > Greetings! I've read the stuff on Tux/X15 in kernel 2.4 with > interest. Apparently there is some enhanced 2.4 kernel tcp interface, > using either TCP_CORK or sendfile(), which I'm not exactly sure, to > effectively get zero-copy tcp/ip networking, and a great performance > boost. I think that if someone had a copy of the source to X15, that > might be the quickest way to find out what's going on. Can't find it > now -- seems to be a commercial product at this point. My question -- > shouldn't the MPI implementations be able to benefit from this?
Yes, the trick is making the incoming buffers (transparently) visible to user processes without going in and out of kernel & copying message buffers. [*] Even the new versions of windows are going to have this so it's a pity that we don't have it in a standard way yet :) Actually, some of the message passing environments on linux do perform similar tricks but any tcp server would benefit from the implementation that you describe, including MPI implementations on linux that employ user space daemons. Any accurate information on fast message passing implementations on linux would be greatly appreciated. Thanks, [*] That was a very bad explanation I guess ;) The truth is that you normally first allocate a buffer in user space. Then if the kernel can get to send that very message without (using the CPU for) copying any array of bytes it's the best. Typically this would involve some interaction with the device driver so the network device driver would have to be smart enough for this fast send routine to work. And doing all that, you shouldn't be causing any more context switches because that would mean latency which is not good for HPC. I'm not very familiar with the linux network protocol stack but from the driver codes I've seen I recognize that this involves a lot of hacking of truly intertwined code. Making network buffers visible through all those layers might require some significant amount of code (and a lot of backspaces). -- Eray Ozkural (exa) <[EMAIL PROTECTED]>, Comp. Sci. Dept., Bilkent University, Ankara www: http://www.cs.bilkent.edu.tr/~erayo GPG public key fingerprint: 360C 852F 88B0 A745 F31B EA0F 7C07 AE16 874D 539C
