It is kind of surprising that random tester can detect the bug with in 3,000,000 cycles while nothing happened on running ruby_fs.py for 200,000,000,000 cycles.
On Tue, December 21, 2010 8:24 pm, Nilay Vaish wrote: > Brad, which protocols work correctly with ruby random tester? > > On Tue, 21 Dec 2010, Beckmann, Brad wrote: > >> Hi Nilay, >> >> If I'm correctly reproducing your problem, I believe I know what the >> issue is. However, before I try to fix it, I want to propose simply >> getting rid of the MESI_CMP_directory. The more and more I look at that >> protocol, the more problems I see. There are several design and logic >> issues in the protocol. Unless someone wants to volunteer to fix them, >> I say we get rid of it as well as all of the protocols not being tested >> by the regression tester. >> >> Now the particular problem that I see causing the deadlock is that that >> L2 cache is drop a PUTX request from the L1 because the L2 is in SS_MB >> state. Thus the L1 remains in M_I state for infinity which of course >> will eventually lead to a deadlock. >> >> Brad >> >> >> -----Original Message----- >> From: [email protected] [mailto:[email protected]] On >> Behalf Of Nilay Vaish >> Sent: Tuesday, December 21, 2010 1:04 PM >> To: [email protected] >> Subject: [m5-dev] Deadlock while running ruby_random_test.py >> >> I am running ALPHA_SE_MESI_CMP_directory with ruby_random_test.py. I >> supply the option -l as 2000. I have pasted the output below. This was >> generated using latest version of m5. >> >> Actually, while testing my own changes to SLICC and protocol files, I >> also observe the dead lock at the 3000001. So I ran the latest version >> and found even that gets stuck. >> >> Is this a known problem? Am I doing some thing wrong? >> >> Thanks >> Nilay >> > _______________________________________________ > m5-dev mailing list > [email protected] > http://m5sim.org/mailman/listinfo/m5-dev > -- Nilay _______________________________________________ m5-dev mailing list [email protected] http://m5sim.org/mailman/listinfo/m5-dev
