[Gems-users] L2 miss comparison of Private and Shared L2 cached systems


Date: Mon, 2 Oct 2006 18:52:32 +0200
From: mehmetderin.harmanci@xxxxxxx
Subject: [Gems-users] L2 miss comparison of Private and Shared L2 cached systems


   Hello all,

      In my current simulations I'm simulating 2 systems to compare them in
   miss performance. The first system is a private L2 cached system with
   4 processors. The second is a shared L2 cached system with 4 processors.
   For pirvate L2 system I used the MOESI-SMP_directory protocol and for
   the shared cached system I used the MSI_MOSI_CMP_directory protocol.
   The results reported bizarrely (for multiple application sets) gives the
   following results: The private L2 cached scheme has a significantly low
   L2 miss rate then the shared L2 cached system! At this point I think
   I'm doing something wrong with the configuration of RUby in the
   simulations. I thoughtthe memory sizes were not comparable in two
   systems but I have not found a contradiction yet. For your information
   I put the memory related parameters below.



    Common to both systems:

                     L1_CACHE_ASSOC:               2
                     L2_CACHE_NUM_SETS_BITS:       7
                     g_DATA_BLOCK_BYTES:         128
                     g_NUM_PROCESSORS:             4
                     g_NUM_L2_BANKS:               4
                     g_MEMORY_SIZE_BYTES: 4294967296


     Specific to each system:



    PRIVATE L2 CACHED SYSTEM                     SHARED L2 CACHED SYSTEM
    (MOESI_SMP_directory prot)                  (MSI_MOSI_CMP_directory prot)

   L2_CACHE_ASSOC:          4                   L2_CACHE_ASSOC:         16
   L2_CACHE_NUM_SETS_BITS: 11                   L2_CACHE_NUM_SETS_BITS: 11

   g_NUM_MEMORIES:          4                   g_NUM_MEMORIES:          1
   g_PROCS_PER_CHIP:        1                   g_PROCS_PER_CHIP:        4



   If I'm not mistaken I need to set L2_CACHE_NUM_SETS_BIT per chip, that is
   for MOESI_SMP protocol I set the size of each private L2 cache (1MB) (data
   block size is 128 Bytes) to the ruby configuration parameters and for the
   MSI_MOSI_CMP I need to set the total L2 cache size (4MB) as a parameter to
   ruby (even if there are multiple L2 banks).

   Is my setting wrong? Can the problem arise form another parameter? Are
   the protocol choices wrong?

   Any help will be greatly greatly appreciated.

   Best Regards,


   Derin Harmanci
[← Prev in Thread] Current Thread [Next in Thread→]
  • [Gems-users] L2 miss comparison of Private and Shared L2 cached systems, mehmetderin . harmanci <=