From: Mike H. <mho...@gr...> - 2004-03-29 23:13:18
|
Karl Rasche wrote: >>>>The bad news. Take a look at NetRecv... We check all of the layers one >>>>by one. I've fixed things in my version so that we only check a layer >>>>if we actually have connections of that type instead of always checking >>>>layers like TCPIP and FILE. However, if we have a TCPIP connection >>>>active, we will call TCPIPRecv. The problem is that if we don't send >>>>data on each tcpip connection (it loops through them all), we sit and >>>>wait for the select and transmit timeouts. >>>> >>>> >>>> >>>> >>>Is the problem that, no matter what, once some data comes in over tcpip, >>>we'll sit in tcpip_read_exact() until all the data has come in? Meanwhile, >>>everyone elses' data is ignored until tcpip is happy? I think i've hit >>>things like this before, but with fighting between tcpip connections, not >>>tcpip vs IB. >>> >>> >>> >>> >>The issue I'm running into seems to be similar. In the case of >>sort-last remote rendering, I want to use a highspeed network for >>compositing and then use tcpip to deliver the final images. So, for >>each frame, there is going to be highspeed traffic(IB) for the >>composite, followed by slow traffic (GigE/100TX) for sending Chromium >>commands on the wire and displaying the final image. However, since we >>check tcpip connections first, I hit the tcpip layer and then sit and >>wait for data to be sent. >> >> > >Ok, I'll try again :) > >So, the problem is waiting for select() to timeout because there isn't any >tcpip goodness coming over the wire yet? > > > Yes. I'll try mucking with the timeout to see if I can get performance back up. I'll also look into what it's going to take to make the network layer not loop through every possible connection... > > >>IB stuff zips along followed by actual tcpip traffic. This drops >>performance achievable over IB from ~700MB/s to about 30MB/s (OUCH!!!). >> >> > >Hmm. one solution would be to trade me some IB for some 100mbps ethernet >;) > I wish things were cheap enough to lend to people. ;-) Actually, any highspeed layer is going to be hosed by this. I thought you guys have a Myrinet cluster there? -Mike |