Linux perf events related to Intel Optane Persistent Memory

[VishalD]

I am trying to measure Intel Optane PMM events using Linux perf tool. The events are unc_m_pmm_rpq_inserts and  unc_m_pmm_rpq_occupancy.all but I don't think these are giving the correct information. 

I modified the STREAM benchmark by replacing the memory allocations to allocate memory on PMM with libmemkind and ran it with 16 threads and array size 1000000000. I also compared it with a normal DRAM implementation and the results are as follows:

The PMM version counters were 

unc_m_pmm_rpq_inserts: 213,724,403 
unc_m_pmm_rpq_occupancy.all: 106,138,605,384

The equivalent DRAM counters for the DRAM version were:

unc_m_rpq_occupancy: 3,362,912,582,581 
unc_m_rpq_inserts: 13,359,982,414 

The numbers are vastly different given that the same application is being run just on different memories with different allocators.

To put this in context, the PMM counters for Intel FIO with job pmemblk.fio running 8 threads were:

unc_m_pmm_rpq_inserts: 253,100,743 
unc_m_pmm_rpq_occupancy.all: 132,347,640,685 

Moreover,

The DRAM version took 30 seconds while the PMM version took 51s which seems reasonable but the triad bandwidth measurement remained same for both at close to 50 Gbps.

So I think there is something wrong with the PMEM counters. Please clarify if I am doing something wrong.

[Victor_G_Intel]

Hello VishalD,

Thank you for posting on the Intel® communities.

We appreciate the contact you made through the forum; however, our support is only here to help with technical requests related to issues with these memories and the tests that you have been doing are something that we cannot validate/support since we are not the developers of the tool you are currently using; therefore, we will suggest, you take your question directly with the developer of the tool to see if they can help you out with it. Additionally, you are always welcome to keep your thread open here in the forum, some of our community peers might be able to help you further.

Regards,

Victor G.

Intel Technical Support Technician

[rkcharles]

I'm attempting to use the Linux perf tool to measure Intel Optane PMM events. Unc m pmm rpq inserts and unc m pmm rpq occupancy are the events. all, but I don't believe they provide accurate information.

I modified the STREAM benchmark by using libmemkind to allocate memory on PMM instead of memory allocations, and I ran it with 16 threads and an array size of 1000000000. I also tested it against a standard DRAM implementation, and the results are as follows:

The PMM version counters were updated.

unc m pmm rpq insertions: 213,724,403 unc m pmm rpq occupancy.

all: 106,138,605,384

For the DRAM version, the equivalent DRAM counters were:

unc m rpq inserts: 13,359,982,414 unc m rpq occupancy: 3,362,912,582,581