[PCM] Adding extra events "MEM_LOAD_UOPS_LLC_MISS_RETIRED.LOCAL_DRAM" and "MEM_LOAD_UOPS_LLC_MISS_RETIRED.REMOTE_DRAM"

[PCM] Adding extra events "MEM_LOAD_UOPS_LLC_MISS_RETIRED.LOCAL_DRAM" and "MEM_LOAD_UOPS_LLC_MISS_RETIRED.REMOTE_DRAM"

I am modifying PCM to monitor other events that are not included in current release. 

More specifically, I am interested in "MEM_LOAD_UOPS_LLC_MISS_RETIRED.LOCAL_DRAM" and "MEM_LOAD_UOPS_LLC_MISS_RETIRED.REMOTE_DRAM".

The "Intel® 64 and IA-32 Architectures Software Developer’s Manual Volume 3 (3A, 3B & 3C): System Programming Guide" comments "Disable BL bypass and direct2core (see MSR 0x3C9)".

I assume that I have to disable BL bypass and direct2core first. Then, I can obtain the values of these events.

However, I did not find anyplace that gives instructions to disable BL bypass and direct2core. 

I would like to know:

1. Do I need to disable BL bypass and direct2core for getting the values of "MEM_LOAD_UOPS_LLC_MISS_RETIRED.LOCAL_DRAM" and "MEM_LOAD_UOPS_LLC_MISS_RETIRED.REMOTE_DRAM"?

2. If the answer is yes to 1, how do I do that?

I appreciate any information and guidance to answer these questions. 

Thank you very much.

Ron C. Chiang

12 posts / 0 new
Last post
For more complete information about compiler optimizations, see our Optimization Notice.

Have you looked at description of MSR 0x3C9 does it contain some info on how to disable BL bypass and direct2core?

Thank you for your reply,

I have checked the description of MSR 0x3C9 in the manual, and searched "BL bypass" and "direct2core" in both the manual and internet.

Unfortunately, I did not find any clear information.

Is there any additional document for MSR?

>>>Is there any additional document for MSR?>>>

I am afraid that beside the SDM documentation there is no freely available more in-depth documentation.

Thank you for your reply,

Does that imply the answer is in some documents that are not freely available?

Also, do I need BIOS support to disable BL bypass and direct2core? If not, can BL bypass and direct2cire be disabled by setting MSR?

Thank you very much!

I suppose that more advanced technical documentation will be available for Bios vendors based upon some kind of NDA agreement.

Unfortunately I do not have an answer to your last question.

Best Reply

Please see this article describing the workaround: http://software.intel.com/en-us/articles/performance-monitoring-on-intel-xeon-processor-e5-family

Intel PCM implements it in PCM::enableJKTWorkaround function.

Roman

Hi Ron

where in the manual is description of MSR 0x3C9? 

Sorry for the late reply and Thanks for the response from iliyapolak and Roman.

iliyapolak,

I can only find Table 18-39 and 35-18 (in the manual Vol.3) mentioned MSR 0x3c9. But, I did not find any explanation there.

Roman, Thank you for the hint. 

My machine is Xeon E5-2620. PCM detects it as "Intel(r) microarchitecture codename Sandy Bridge-EP/Jaketown"

After seeing your reply and trace the code again, I found that workaround is already enabled. So, I started to debug my modification.

The original pcm.x monitors 4 events on this machine. My first attempt was extending "coreEventDesc" in cpucounters.cpp. For example, I add the following code for one extra event. 

"coreEventDesc[4].event_number = MEM_LOAD_UOPS_LLC_MISS_RETIRED_LOCAL_DRAM_EVTNR;

coreEventDesc[4].umask_value = MEM_LOAD_UOPS_LLC_MISS_RETIRED_LOCAL_DRAM_UMASK;"

I also changed corresponding numbers, e.g, core_gen_counter_num_used = 5; //It was 4 before the modification.

However, I never get a valid number from this implementation. 

Then, I tried another way today. I change one of the 4 originally mornitored events to the one I want, instead of adding extra one to "coreEventDesc".

It works. Now I can collect other events.

I am happy to have this problem solved. But, I have one more question. pcm.x shows that "Number of core PMU generic (programmable) counters: 8". So, I thought that I have 8 counters to use and started my first attempt. However, it seems there are only 4 available?

Ron

The Sandy Bridge core has 8 programmable performance counters, but these are split into two groups of four when HyperThreading is enabled.

John D. McCalpin, PhD
"Dr. Bandwidth"

Thank you, John,

I did disable HyperThreading on this machine.

Hi,

when Hyperthreading is disabled there should be 8 programmable counters in hardware on your system. 4 hardware programmable counters are available if Hyperthreading is enabled. Intel PCM currently supports 4 programmable counters. We might address this in a future version of Intel PCM.

Roman

Leave a Comment

Please sign in to add a comment. Not a member? Join today