Hi Kypros, The network tester generates (InjectionRate * num-dirs * sim-cycles) number of flits, and these are sent to the NetworkInterface (NI) in Garnet. The NI now tries to inject them into the network. That depends on whether there are free VCs at the input port of the connected router, which in turn depends on network congestion etc. The "total flits injected" stat is only incremented upon a successful injection into the network. Other generated flits wait at the NI.
You specified an injection rate of 1 flit/cycle. However, for a k x k mesh, the network theoretically saturates at 4/k flits/node/cycle = 0.5 flit/node/cycle for Uniform Random Traffic (i.e. this is the maximum number of flits that can be injected to keep all links occupied). In reality, it saturates even earlier because we use simple XY routing, and switch allocation is never perfect etc so not all links get used efficiently … In your case, the network seems to have saturated at (2477428/64/100000) = 0.38 flits/node/cycle which seems correct. Post saturation, the flits that are waiting at the NI will start having huge queueing latencies. This should be reflected in your latency stats. cheers, Tushar On Aug 2, 2012, at 8:30 AM, Kypros Chrysanthou wrote: > Hello, > > I noticed that even though the network tester generates round about > the expected number of packets, based on the number of flits > ( InjectionRate * num-dirs * sim-cycles ) > e.g 64 dirs, 1.0 injectionRate (measured in flits/node/cycle) > and 100,000 sim-cycles the result should be about 6,400,00 flits, > the injected flits though are > > from ruby.stats > Total flits injected = 2477428 > Total flits received = 2474033 > > Why does this occur? > > Thanks, > Kypros > _______________________________________________ > gem5-users mailing list > [email protected] > http://m5sim.org/cgi-bin/mailman/listinfo/gem5-users
_______________________________________________ gem5-users mailing list [email protected] http://m5sim.org/cgi-bin/mailman/listinfo/gem5-users
