From: David J. <jo...@mc...> - 2003-09-26 22:03:15
|
Yeah, I made all appropriate changes to the server and app node names, as well as changing the protocol to 'gm' for AddServer. I did try running the simplemural config locally using gm and I got pretty miserable performance for atlantis (~1.2fps) with an MTU of 1024*1024 using two crservers. This is a fairly new system I'm using too; there could be some general problems with myrinet. -Dave On Thu, 25 Sep 2003, Mike Houston wrote: > crdemo.conf and simplemural.conf works on my end. This is with the top > of tree. > > Did you remember to modify the conf file to actually name the > application and server systems? ie for crdemo.conf server_node = > CRNetworkNode( 'system1') and client_node = CRApplicationNode( 'system2' > ). Those changes as well as replacing all 'tcpip' with 'gm' gets > Myrinet working on my end. > > A few notes on our GM installation: We are running version 1.6.4 and > use a static mapping for the nodes. > > -Mike > > David Jones wrote: > > I tried it again with a cvs checkout from yesterday. crdemo.conf still > > hangs. I threw in a few crDebugs and it seems like it's getting stuck in > > an infinite loop in both packspu_context.c and tilesortspu_context.c. > > > > ... > > > > else { > > while (writeback) <---- this never gets changed. > > crNetRecv(); > > > > if (pack_spu.swap) { > > serverCtx = (GLint) SWAP32(serverCtx); > > } > > if (serverCtx < 0) { > > #ifdef CHROMIUM_THREADSAFE > > crUnlockMutex(&_PackMutex); > > #endif > > crWarning("Failure in packspu_CreateContext"); > > return -1; /* failed */ > > } > > } > > > > ... > > > > It seems like crGmRecv() is getting called a bunch of times, but nothing > > ever happens for it to break out of the loop. > > > > -Dave > > > > > > > > On Wed, 24 Sep 2003, Mike Houston wrote: > > > > > >>You need to setup the environment variable for CRMOTHERSHIP to point to > >>the system on which the mothership is running or feed that flag to the > >>crserver and crappfaker. > >> > >>GM is *almost* fully operational. The only thing that doesn't work > >>correctly at the moment is setting larger MTUs for SPUs high in the > >>chain than MTUs of the SPUs lower in the chain. For example, if you can > >>set a different MTU for the binaryswap SPU that that of the system. > >>However, if you set this MTU larger than the pack SPU's MTU (it uses the > >>system default ie cr.MTU(x)), then the connections will hang... > >> > >>All of the simple conf files in the tree work on my end. I haven't > >>tested CRUT yet over GM yet... > >> > >>-Mike > >> > >> > >>Greg Humphreys wrote: > >> > >> > >>>Hi Dave. Currently, servers don't automatically quit on non-tcp networks. > >>>I think that's on the short TODO list. > >>> > >>>There was some flakiness that was introduced into the network layer a couple > >>>of months back, and since not enough people use high-speed networks it > >>>wasn't caught early. You *might* be seeing some of that. Mike Houston is > >>>working hard on getting everything back up to par, so he should weigh in on > >>>whether GM works for even simple stuff. > >>> > >>>Mike? > >>> > >>>-- > >>>Greg Humphreys, Assistant Professor > >>>Department of Computer Science, University of Virginia > >>>hu...@cs... > >>> > >>> > >>> > >>> > >>>>-----Original Message----- > >>>>From: chr...@li... > >>>>[mailto:chr...@li...] On Behalf > >>>>Of David Jones > >>>>Sent: Wednesday, September 24, 2003 2:01 PM > >>>>To: chr...@li... > >>>>Subject: [Chromium-users] myrinet and 1.4 > >>>> > >>>> > >>>>I decided to try out chromium with myrinet support. I've > >>>>never dealt with > >>>>gm stuff before, so this is all pretty new for me. I tried the > >>>>crdemo.conf with 'gm' used as the protocol for > >>>>client_spu.AddServer. I > >>>>left the MTU as 1024*1024. > >>>> > >>>>Running everything on the localhost works fine, except the crserver > >>>>doesn't die when I quit atlantis. I ran into issues when I > >>>>ran the server node on a different machine. Both the crappfaker and > >>>>crserver seem to hang. The final output from crserver is: > >>>> > >>>>CR Warning(tg-v007:24404): GM: write pool: bufsize=2056, 4080 > >>>>buffers, 8388480 bytes > >>>>CR Warning(tg-v007:24404): crGmAccept is being called -- > >>>>brokering the connection through the mothership!. > >>>>CR Warning(tg-v007:24404): Total output dimensions = (500, 500) > >>>> > >>>>The crappfaker shows: > >>>> > >>>>CR Warning(tg-viz-login1:26736): GM: write pool: > >>>>bufsize=2056, 4080 buffers, 8388480 bytes > >>>>CR Warning(tg-viz-login1:26736): crGmDoConnect is being called -- > >>>>brokering the connection through the mothership!. > >>>>CR Warning(tg-viz-login1:26736): Couldn't find the > >>>>CRMOTHERSHIP environment variable, defaulting to localhost > >>>> > >>>>I tried messing with the MTU values, but that didn't get me > >>>>anywhere. I'm > >>>>using the gm-2.0 API. Can any myrinet experts suggest anything? > >>>> > >>>>Thanks much > >>>>Dave J. > >>>> > >>>> > >>>> > >>>>------------------------------------------------------- > >>>>This sf.net email is sponsored by:ThinkGeek > >>>>Welcome to geek heaven. > >>>>http://thinkgeek.com/sf > >>>>_______________________________________________ > >>>>Chromium-users mailing list > >>>>Chr...@li... > >>>>https://lists.sourceforge.net/lists/listinfo/chromium-users > >>>> > >>>> > >>>> > >>> > >>> > >>> > >>>------------------------------------------------------- > >>>This sf.net email is sponsored by:ThinkGeek > >>>Welcome to geek heaven. > >>>http://thinkgeek.com/sf > >>>_______________________________________________ > >>>Chromium-users mailing list > >>>Chr...@li... > >>>https://lists.sourceforge.net/lists/listinfo/chromium-users > >>> > >>> > >> > >> > > > > > > > > ------------------------------------------------------- > > This sf.net email is sponsored by:ThinkGeek > > Welcome to geek heaven. > > http://thinkgeek.com/sf > > _______________________________________________ > > Chromium-users mailing list > > Chr...@li... > > https://lists.sourceforge.net/lists/listinfo/chromium-users > > |