A professor I was in touch with is trying to do this with a highly modified M5 simulator. They can do up to 256 cores with a directory based coherence protocol. I know from experimenting with their simulator that they can only successfully do VERY simple core models when scaling to that many cores (i.e. in order, 1 CPI, non pipelined cores) in M5 system emulation mode. Even then they were burning through like ~50GB of memory just to simulate the directory protocols. I would be very impressed if you were able to do this scale with MARSS, but I don't know if I'd bet on it.
I tried to see if they've published/released anything, but I don't see anything yet. If that sounds interesting I can ask to see what their thoughts are on a public release. Are you trying to simulate shared memory (like many-core CMP) or distributed memory (like super computer MPI nodes)? One thing we've been a part of is the SST framework ( http://www.cs.sandia.gov/sst/ ) at Sandia National Labs -- it's built to be scalable across MPI nodes -- so you can instantiate a ton of CPU instances and spread them out over a giant cluster. This project is kind of in a state of flux and a lot of the components are not really up to snuff at the moment (the DRAMSim component works great though ;) Not sure if any of that is helpful. -Paul On Mon, Sep 27, 2010 at 4:45 PM, Kevin Macdonald <[email protected]>wrote: > I am currently looking into using Marss as my simulation platform, and I > have a few general questions. > > 1. I would like to be able to simulate a large numbers of cores (128 or > potentially more). I noticed the performance graph on the main website only > indicated up to 8 cores, do you think Marss can handle 128 reasonably? Can > Marss utilize a bigger/better computer with more processors and cores to > make a single simulation go faster, or will the only way to increase > throughput be to run multiple Marss simulations? > > 2. I'm interested in expanding on the interconnect portion of Marss. How > possible do you guys think it would be for me to incorporate a pre-existing > standalone network simulator into Marss to perform the interconnect portion > of the simulation (I was thinking of booksim). This would also pretty much > require that I implement directory based protocols for cache coherence, I'm > assuming this shouldn't be too hard to add? > > 3. Is there already ongoing work of the nature that I described above? I > don't necessarily want to attempt to duplicate something that others are > already hard at work on. > > Thanks! > > -Kevin Macdonald > > _______________________________________________ > http://www.marss86.org > Marss86-Devel mailing list > [email protected] > https://www.cs.binghamton.edu/mailman/listinfo/marss86-devel > >
_______________________________________________ http://www.marss86.org Marss86-Devel mailing list [email protected] https://www.cs.binghamton.edu/mailman/listinfo/marss86-devel
