> > Message: 1 > Date: Mon, 16 Feb 2015 19:20:21 -0500 > From: Poovaiah MP <[email protected]> > To: [email protected] > Subject: [marss86-devel] Dependence of IPC on cache latency > Message-ID: > <CAPu_Uo+JeEHYMuhJoGh6McP69MjCjN=A7_3oxMa8ZX9+Lo-f= > [email protected]> > Content-Type: text/plain; charset="utf-8" > > Hi everyone, > > I'm trying to understand the dependency of IPC on (shared) L2 cache > latency. I would expect naively that IPC would reduce when L2 cache latency > is increased. However when I change the L2 latency to 5, 10, and 20 cycles, > the IPC reported is about 4.53, 4.37, and 4.51. This is counter intuitive > for me. Is there something that I'm missing here? > > I appreciate any help in this regard. > > Thanks and regards, > *Poovaiah* > *Circuits and systems laboratory,* > *University of Pittsburgh* > > Addendum: > > Related thread: > http://thread.gmane.org/gmane.comp.emulators.marss86/319/focus=321 > > Details of my setup: > I'm running a spec benchmark on c=4 machine for 100million instructions. > The IPC is grepped from the stats file and summed together. >
I've seen similar problem before, and for me the problem was simulation variability. I suggest you run some more tests to find out what's going on: (1) Use the same latency, run your experiment multiple times, and see if the IPCs are similar. Some benchmarks have very high IPC variation. (2) Run more instructions. (3) Use a smaller L2. My concern for simulating small number of instructions + large L2 is that most of your cache misses are cold misses, so L2 latency will be less of a performance deciding factor. Mu-Tien -------------- next part -------------- > An HTML attachment was scrubbed... > URL: < > https://www.cs.binghamton.edu/mailman/private/marss86-devel/attachments/20150216/d5d36d22/attachment-0001.html > > > > ------------------------------ > > _______________________________________________ > http://www.marss86.org > Marss86-Devel mailing list > [email protected] > https://www.cs.binghamton.edu/mailman/listinfo/marss86-devel > > > End of Marss86-Devel Digest, Vol 60, Issue 2 > ******************************************** >
_______________________________________________ http://www.marss86.org Marss86-Devel mailing list [email protected] https://www.cs.binghamton.edu/mailman/listinfo/marss86-devel
