Hmmm the immediate thought that comes into mind when you change the routing algorithm is deadlocks … It should also deadlock with PARSEC but perhaps the network is not very congested with PARSEC which is why the deadlock scenario doesn't occur … If you place debug flags in the SWallocator you should be able to figure out why flits are not going out … My guess is that you will notice a cyclic dependency in buffers…
- Tushar On Apr 4, 2012, at 4:16 AM, Dean Michael Ancajas wrote: > Hi Tushar, > Thanks for the explanation. Examining the TESTER code now I understand it > more. > > I do have some other follow up questions. What could be a primary reason that > a network-testing only simulation could stall? Suddenly it just stops routing > packets (I have put printf/cout statements in routeCompute and they are not > executed. ) Also, if I use debugger embedded with the tester, the EventQueue > is still running and calling triggerAllEvents() and the cycle count is also > increasing so its not like the simulation itself has stopped. For some > reason, the whole network just stops routing. The only thing I have changed > from the vanilla gems-garnet config is the routing algorithm, instead of > using the XY routing, I chose links that are pre-generated offline. What's > interesting though is that the same routing algorithm is perfectly working > when running full-system simulation of simics+gems+garnet (using PARSEC > benchmarks) config so this got me a bit puzzled for a while now. > > thanks, > dean > > > > On Tue, Apr 3, 2012 at 5:38 PM, Tushar Krishna <tus...@csail.mit.edu> wrote: > Hi Dean, > The number of vnets that get traffic ultimately depend on the coherence > protocol. When you ran Simics, the coherence protocol you built will > determine how many vnets are active and what kinds of traffic they carry. > > If you are using the GEMS network tester, the traffic injector code in > Tester.C might be hard coded to only inject traffic in vnet 0. If I remember > correctly, it injects both control (1-flit) and data (5-flit typically) > packets. You can edit it if you want. > > If you are using the gem5 network tester, details of which can be found here > (http://www.m5sim.org/Networktest), it will inject traffic in three of the > vnets as the wiki explains. > > - Tushar > > > On Apr 3, 2012, at 7:03 PM, Dean Michael Ancajas wrote: > >> Hi, >> My question is regarding the garnet network only tester from this: >> https://lists.cs.wisc.edu/archive/gems-users/2009-October/msg00030.shtml >> >> I have made it run but I am wondering if there is a special reason why only >> the first virtual network is used. If I run under simics I am getting >> traffic from 16 out of 20 Virtual channels. >> >> >> Average VC Load [0] = 7.23417 flits/cycle >> Average VC Load [1] = 7.20467 flits/cycle >> Average VC Load [2] = 7.14563 flits/cycle >> Average VC Load [3] = 7.05776 flits/cycle >> Average VC Load [4] = 0 flits/cycle >> Average VC Load [5] = 0 flits/cycle >> Average VC Load [6] = 0 flits/cycle >> Average VC Load [7] = 0 flits/cycle >> Average VC Load [8] = 0 flits/cycle >> Average VC Load [9] = 0 flits/cycle >> Average VC Load [10] = 0 flits/cycle >> Average VC Load [11] = 0 flits/cycle >> Average VC Load [12] = 0 flits/cycle >> Average VC Load [13] = 0 flits/cycle >> Average VC Load [14] = 0 flits/cycle >> Average VC Load [15] = 0 flits/cycle >> Average VC Load [16] = 0 flits/cycle >> Average VC Load [17] = 0 flits/cycle >> Average VC Load [18] = 0 flits/cycle >> Average VC Load [19] = 0 flits/cycle >> _______________________________________________ >> gem5-users mailing list >> gem5-users@gem5.org >> http://m5sim.org/cgi-bin/mailman/listinfo/gem5-users > > > _______________________________________________ > gem5-users mailing list > gem5-users@gem5.org > http://m5sim.org/cgi-bin/mailman/listinfo/gem5-users > > _______________________________________________ > gem5-users mailing list > gem5-users@gem5.org > http://m5sim.org/cgi-bin/mailman/listinfo/gem5-users
_______________________________________________ gem5-users mailing list gem5-users@gem5.org http://m5sim.org/cgi-bin/mailman/listinfo/gem5-users
