[Ptools-perfapi] Count L3 cache misses on an AMD Opteron 6168
eranian at googlemail.com
Fri Sep 17 11:37:54 EDT 2010
On Fri, Sep 17, 2010 at 5:21 PM, Vince Weaver <vweaver1 at eecs.utk.edu> wrote:
> On Fri, 17 Sep 2010, Juri Lelli wrote:
>> I have no problems to read L1 and L2 cache misses, but I'd like to read L3 cache misses too. papi_avail
>> | grep L3 returns:
> L3 is tricky to measure, as the L3 is a shared resource used by multiple
> cores. That's why it doesn't have a PAPI predefined event available.
> You can still measure things though, but you'll have to use a native
> from util/papi_native_avail on an Istanbul chip
> 0x40000079 L3_CACHE_MISSES | L3 Cache Misses |
> 40000479 :READ_BLOCK_EXCLUSIVE | Read Block Exclusive (Data cache read) |
> 40000879 :READ_BLOCK_SHARED | Read Block Shared (Instruction cache read) |
> 40001079 :READ_BLOCK_MODIFY | Read Block Modify |
> 40002079 :CORE_0_SELECT | Core 0 Select |
> 40004079 :CORE_1_SELECT | Core 1 Select |
> 40008079 :CORE_2_SELECT | Core 2 Select |
> 40010079 :CORE_3_SELECT | Core 3 Select |
> 40020079 :CORE_4_SELECT | Core 4 Select |
> 40040079 :CORE_5_SELECT | Core 5 Select |
> 40080079 :ANY_CORE | Any core |
> 40100079 :ALL | All sub-events selected |
Careful when using this event. There is an erratum published about many
Northbridge events with a CORE_SELECT type of umasks. They don't work.
The only valid choice you have is ANY_CORE.
I have updated libpfm-3.10 and GIT libpfm4 to take this into account.
> Ptools-perfapi mailing list
> Ptools-perfapi at eecs.utk.edu
More information about the Ptools-perfapi