Hi Akanksha, Thanks a lot for the reply. We are simulating 500M instructions now. We will try to simulate the server and client in different MARSS instances to see whether that gives better results.
FYI, if you are using the directions to run memcached from cloudsuite page, the flags for running memcached are given wrong there. Specifically, the command to run memcached should be ./memcached -m <amount of memory in MB> ... instead of ./memcached -D <amount of memory in MB> .... The -D flag actually turns on fine grained statistics collection in memcached which consumes most of cycles in memcached. You can verify this by doing ./memcached -h. When simulating with that configuration we also got miss rates of ~ 10%, but it dropped when we used -m. -- Subhasis On Tue, Oct 22, 2013 at 8:25 PM, Akanksha Jain <[email protected]>wrote: > Subhasis, > > How many instructions are you simulating? It might be a good idea to > run longer simulations to get representative miss rates because Cloud > Suite benchmarks are fairly I/O intensive. On simulating a total of > 250 million instructions with a single threaded server, my experiments > show a miss rate of 11% for memcached. > > Also, are you running the client natively? Given that Marss is orders > of magintude slower than native execution, you might try running the > client in a separate instance of Marss to match the client's request > rate to the server's processing rate. > > I did not attend the MICRO 2012 tutorial, so I am curious what changes > were suggested to run Cloud Suite with Marss. In our experience so > far, it takes an unrealistic amount of simulation time to warm up the > caches. As a result, we observe very little memory activity when > simulating 1-2 billion instructions (memcached was an exception). Our > guess is that due to the JVM and I/O overheads, it takes Marss a > really really long time to process enough data to warm up the caches. > Also fast-forwarding through the initialization phase in a multi-core > setting seems to hang a lot of times. Does anybody have any thoughts > on these issues? > > On Tue, Oct 22, 2013 at 9:52 PM, Subhasis Das <[email protected]> > wrote: > > We are trying to study some datacenter applications and are trying to get > > cloudsuite to work under Marssx86. > > > > We have some of the benchmarks from cloudsuite running under Marssx86. > > However the simulated cache miss rate results seem to be very different > from > > cache miss rates obtained from native hardware using perf (e.g., for > > memcached the native L1 miss rates are ~ 15%, whereas we observe only ~ > 3%). > > The setup we are using for simulation is to have the server workload > running > > in a single threaded simulated machine while we run the load generators > from > > outside and we use qemu's user mode networking. > > > > You mentioned in your MICRO 2012 tutorial that the work to simulate > > cloudsuite under Marssx86 is ongoing. Is that work done? If yes, is it > > possible for us to gain access to the necessary steps for simulating > > cloudsuite on Marssx86, and/or the disk images used by you? > > > > -- > > Thanks, > > Subhasis > > > > _______________________________________________ > > http://www.marss86.org > > Marss86-Devel mailing list > > [email protected] > > https://www.cs.binghamton.edu/mailman/listinfo/marss86-devel > > > > > > -- > Thanks, > Akanksha Jain > PhD Candidate > Department of Computer Science > University of Texas at Austin >
_______________________________________________ http://www.marss86.org Marss86-Devel mailing list [email protected] https://www.cs.binghamton.edu/mailman/listinfo/marss86-devel
