Profiling memory accesses - which counters overlap with which?

[Adrian_C_1]

Hi!

I have some Perl scripts that implement some of the top-level vtune profiling for sandy bridge (and SB Xeon.)

Noticably:

* % cycles spent LLC miss: ~ 53%

* % cycles spent doing DTLB walks: ~ 10%

* % cycles spent accessing data modified by another core: ~ 3.5%

My question: when doing DTLB walks (DTLB_LOAD_MISSES.WALK_DURATION) and cycles spent hit'ing on data modified by another core (MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_HITM) - do these overlap in any way with the LLC_MISS (ie memory access - MEM_LOAD_UOPS_MISC_RETIRED.LLC_MISS) ? Partially overlap? Don't overlap at all?

I'm especially interested in whether the DTLB walks count towards MEM_LOAD_UOPS_MISC_RETIRED.LLC_MISS.

Thanks!

-adrian

[Patrick_F_Intel1]

Hello Adrian,

I'd have to see the full equation for the metric. Can you post them? Even with the equation it might be hard to tell.

Pat

[Adrian_C_1]

Hi,

The code is at github.com/erikarn/hwpmc/ . The equations match what's in this PDF:

https://software.intel.com/sites/landingpage/legacy/pdfs/Using_Intel_VTune_Amplifier_XE_on_2nd_Gen_Intel_Core_Family.pdf

Thanks!

-adrian

[Adrian_C_1]

Hi,

The code is at github.com/erikarn/hwpmc/ . The equations match what's in this PDF:

https://software.intel.com/sites/landingpage/legacy/pdfs/Using_Intel_VTune_Amplifier_XE_on_2nd_Gen_Intel_Core_Family.pdf

Thanks!

-adrian

[Patrick_F_Intel1]

A 42 pages of images, unsearchable document. nice.

[Adrian_C_1]

The PDF display for me has a bunch of images but the formulas are in text, so they're searchable.

[Adrian_C_1]

bump?