Dear Alexandros, a) The directory and the L2 tag array are looked up in parallel and the L2 data array is looked up after. So the L2 hit time is: MAX(dir_lat, tag_lat) + data_lat. b) There is a network topology file which sets various interconnection delays. The name of the topology file is specified in the configuration file (e.g., "16x3-mesh.topology") and the file itself is located in the simulator directory (e.g., simulators/CMP.L2SharedNUCA.OoO/16x3-mesh.topology). Per hop latency is equal to the channel latency plus a fixed 2-cycle router latency.
Regards, Mohammad On Jun 12, 2012, at 10:21 AM, aledaglis wrote: > Hello everybody, > > I have been experimenting with some new designs on the CMP.L2SharedNUCA.OoO > simulator and I am now on the process of evaluating my results. > > I am not sure whether I got some L2 access timing parameters right, so I > would really appreciate it if you could help me get these things straight. > > > a) The hit-time for a given L2 slice is given to the simulator through the > user-postload.simics spec file. Could you please confirm whether > the hit-time is evaluated as the sum of the following parameters: > > flexus.set "-L2:dir_lat" "4" # "Total > latency of directory lookup" (DirLatency) > flexus.set "-L2:tag_lat" "4" # "Total > latency of tag array lookup" (TagLatency) > flexus.set "-L2:data_lat" "8" # "Total > latency of data array lookup" (DataLatency) > > ... so a hit for this configuration would take 4+4+8 = 16 cycles. > > > b) Could you give me a hint about the cost, in cycles, for each hop in the > tiled architecture? I guess that for a remote L2 hit, the overall > cost would be > > <L2 hit time> + <hop time> * <number of hops> > > Is the <hop time> hard-coded or can it be also defined through the config > file as a parameter of the interconnection? > > Thank you in advance! > Alexandros Daglis
