Answers inline. On Sep 17, 2012, at 12:42 PM, Pavan Poluri wrote:
> Hello, > > I have a question on the ni_flit_size parameter in the file > garnet/BaseGarnetNetwork.py. From the documentation I understood that > ni_flit_size specifies the flit size in bytes. The default value is 16 bytes. > In the documentation, it says that this results in a control message fitting > in 1 flit and data message fitting in 5 flits. So this means that the control > message is 16 bytes and the data message is 80 bytes. The following are the > two questions I have: > > 1. Lets say if I change the ni_flit_size to 8 bytes, would it automatically > translate to a control message that fits in 2 flits and data message fitting > in 10 flits? > Yes. Take a look at NetworkInterface_d.cc where number of flits are calculated. > 2. Are the sizes of control message (16 bytes) and data message (80 bytes) > fixed? Is it possible to modify their sizes? > Take a look at network/Network.py/cc > Thanks for your time. > > Thanks, > Pavan > > On Wed, Sep 12, 2012 at 2:54 PM, Pavan Poluri <poluripa...@gmail.com> wrote: > Thanks Tushar. > > Thanks, > Pavan > > > On Wed, Sep 12, 2012 at 12:23 PM, Tushar Krishna <tus...@csail.mit.edu> wrote: > If you divide the total flits by total cycles by 16, you can see that the > injection rate is only 0.0009 flits/cycle/node. Hence your power is so low. > The total network energy might be an alternate metric that you might want to > consider instead of power to remove cycles from the picture. > Take a look at src/mem/ruby/network/orion/NetworkPower.cc where the energy > and power calculations are done. > For a relative comparison, the numbers from Orion might work for you... > You could compare the energy numbers for each component from Orion and DSENT > if you want to see how much they differ. > > The cache sizes are in configs/common/Options.py > > > On 09/12/2012 03:26 PM, Pavan Poluri wrote: >> >> Hi Tushar, >> >> The simulation ran for 5,400,912,679 cycles. How do I reduce the cache >> sizes? Which source files do I need to modify? >> >> I was also looking into DSENT tool. To the extent I understood, the current >> version of DSENT does not model the power of the Virtual Channel Allocation >> stage. It only models the power for buffer, crossbar, switch allocator and >> clock. I really need to calculate the power of the Virtual Channel >> Allocation stage. >> >> Thanks for your help. >> >> Thanks, >> Pavan >> >> On Wed, Sep 12, 2012 at 10:51 AM, Tushar Krishna <tus...@csail.mit.edu> >> wrote: >> Hi Pavan, >> There are two issues here. >> One, as Mitch pointed out, is that Orion is not entirely accurate. >> I would suggest computing activity counts from garnet and feeding them to >> DSENT. >> >> However, I have a feeling you will see a similar phenomenon (dynamic power >> >> leakage power) even with DSENT. >> How many cycles did your simulation run for? >> For full system runs in gem5, the network activity is typically very low >> (since network gets flits only on cache misses). >> As a result your dynamic power is very low. >> Network activity can be increased by reducing cache sizes. >> >> cheers, >> Tushar >> >> >> On Sep 12, 2012, at 1:43 PM, Pavan Poluri wrote: >> >>> Hi, >>> >>> Thanks a lot for your detailed reply. >>> >>> Thanks, >>> Pavan >>> >>> On Wed, Sep 12, 2012 at 10:32 AM, Mitch Hayenga >>> <mitch.hayenga+g...@gmail.com> wrote: >>> Hi, >>> >>> I wouldn't trust the power model. Garnet is based on Orion, which in the >>> last year a few papers have shown to be quite inaccurate (mostly because >>> its internal model doesn't scale some technology parameters properly). >>> >>> More Information: >>> 1. Peh's group recently announced a more accurate power modeling tool >>> called DSENT (https://sites.google.com/site/mitdsent/). In their paper >>> they highlight many issues with Orion and (at the 45nm node) find it >>> capable of being off by ~10x in power. >>> >>> 2. I published a WDDD paper on Orion showing my own brief investigation >>> into why its power/area numbers seemed disconnected with reality. >>> (http://www.ece.wisc.edu/~hayenga/papers/wddd2012_hayenga.pdf) >>> >>> Hope this helps. Maybe the version of Orion integrated with Ruby/gem5 has >>> received some updates, but unless you've heard otherwise, I wouldn't trust >>> it. >>> >>> >>> On Wed, Sep 12, 2012 at 12:30 PM, Mitch Hayenga <mitch.haye...@gmail.com> >>> wrote: >>> Hi, >>> >>> I wouldn't trust the power model. Garnet is based on Orion, which in the >>> last year a few papers have shown to be quite inaccurate (mostly because >>> its internal model doesn't scale some technology parameters properly). >>> >>> More Information: >>> 1. Peh's group recently announced a more accurate power modeling tool >>> called DSENT (https://sites.google.com/site/mitdsent/). In their paper >>> they highlight many issues with Orion and (at the 45nm node) find it >>> capable of being off by ~10x in power. >>> >>> 2. I published a WDDD paper on Orion showing my own brief investigation >>> into why its power/area numbers seemed disconnected with reality. >>> (http://www.ece.wisc.edu/~hayenga/papers/wddd2012_hayenga.pdf) >>> >>> Hope this helps. Maybe the version of Orion integrated with Ruby/gem5 has >>> received some updates, but unless you've heard otherwise, I wouldn't trust >>> it. >>> >>> >>> On Wed, Sep 12, 2012 at 12:07 PM, Pavan Poluri <poluripa...@gmail.com> >>> wrote: >>> Hello, >>> >>> I have executed the Blackscholes application of the PARSEC benchmark suite >>> with 16 threads on the input file set (in_4.txt) with a full system >>> simulation with 16 cores, 16 L2 caches and 16 directories on a mesh >>> topology with 4 rows. I have used the MOESI_CMP_directory protocol. The >>> technology used is 90nm with a clock frequency of 1GHz and operating >>> voltage VDD of 1.2V. I was going through the power statistics in the >>> ruby.stats file. The following are the power numbers from the simulation. >>> >>> Router Dynamic Power = 0.00710691 W => 0.4441 mW per router >>> Router Static Power = 0.452366 W => 28.272 mW per router >>> Router Clock Power = 0.541901 W >>> >>> I am confused with these power numbers. The dynamic power is very very less >>> compared to the static power. I do not understand why the dynamic power is >>> so low even when the simulation resulted in the injection of 75,899,868 >>> flits and the successful reception of 75,899,865 flits. Am I doing >>> something wrong with the simulation? Do I need to set some parameters for >>> the power calculations? >>> >>> Thanks for your time. >>> >>> Thanks, >>> Pavan >>> >>> _______________________________________________ >>> gem5-users mailing list >>> gem5-users@gem5.org >>> http://m5sim.org/cgi-bin/mailman/listinfo/gem5-users >>> >>> >>> >>> -- >>> Mitch Hayenga >>> mitch.haye...@gmail.com >>> >>> >>> _______________________________________________ >>> 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 > > _______________________________________________ > 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
