1 Reply Latest reply on Apr 3, 2012 11:03 AM by fswehosky

    Data Cache Misses on Bobcat

    ebaron

      I'm trying to get an accurate reading of L1 Data cache misses on my E-350 processor. The document "Basic Performance Measurements for AMD Athlon™ 64, AMD Opteron™ and AMD Phenom™ Processors" has been useful in interpreting results from OProfile. The document states that the event 0x41, DATA_CACHE_MISSES, can be inaccurate. The more precise method is adding the results from events 0x42, DC_refills_L2, and 0x43, DC_refills_sys. In the 14h BKDG, event 0x42 is now called DATA_CACHE_REFILLS_FROM_L2_OR_NORTHBRIDGE, and 0x43 is DATA_CACHE_REFILLS_FROM_NORTHBRIDGE.

       

      Does event 0x42 now include the misses measured by 0x43? That is, should I only use results from 0x43, or the sum of 0x42 and 0x43?

       

      Thanks,

      Elliott

        • Re: Data Cache Misses on Bobcat

          Hi Elliott,

           

          From my reading of the BKDG, the count from event 0x42 includes data cache refills satisfied from the L2 cache and the Northbridge.  0x43 counts are only from the Northbridge.  I'll point out that the Family 14h BKDG doesn't mention any of the family 10h streaming store issues with 0x41, so it might be useful to use the 0x41 event. 

           

          In the paper you referenced for a family 10h system, the unit mask for 0x42 had bit 0 as 0, so adding the 0x42 and 0x43 would work.  In the family 14h BKDG, the unit mask for event 0x42 has changed, so you can no longer mask out the Northbridge (0x43) events.  So, it looks like the count for 0x42 on your system would include the count for 0x43.

           

          If you want misses, I suggest 0x41.  If you want data cache refills just from the L2 cache, I recommend the count of 0x42 - 0x43.  If you want the data cache refills from the Northbridge, use event 0x43.

           

          Does this help?

          Thanks,

          -=Frank

          1 of 1 people found this helpful