[BUG] Incorrect Memory bandwidth measurement on AMD Threadripper 2950x (Zen+)

[badtst]

Describe the bug When measuring memory bandwidth using likwid-perfctr in stethoscope mode and group MEM, reported memory bandwidth is ~ half of the memory bandwidth reported by the benchmark and AMDµProf measurements.

Looks like 2950x has 4 memory channels, two per die (2 for CCD-0 and 2 for CDD-1).

Threadripper 2950x with 4x16GB RAM and all 4 DRAM channel being used (Output of dmidecode --type 17 provided).

dmidecode17.txt

To Reproduce

• OS : Ubuntu 20.04.5 LTS
• Build Likwid from source Version 5.2.2 (commit: 233ab943543480cd46058b34616c174198ba0459)
• Run a memory intensive benchmark i.e. stress-ng --memcpy 0
• Measure memory with Likwid
• Compare output with AMDuProfPcm -m memory -a -q -d 2000 for example, which is approximately double of what's reported by Likwid.

AMDuProf.txt

To Reproduce with a LIKWID command Output of the command with -V 3 added to the command:

• likwid-perfctr -f -c 0-31 -g MEM -S 2s
• likwid-topology

perfctr.txt topology.txt

Additional context Processor Programming Reference (PPR) for AMD Family 17h Models 01h,08h, Revision B2 Processors p161 writes that the measurement of transferred Data bytes is per Node which they define as : A node, is an integrated circuit device that includes one to 8 cores (one or two Core Complexes). i.e. CDD ? Which is detected by likwid-topology as a different die and not a different socket.

[TomTheBear]

Thanks for reporting. Please check the mentioned PR.

[badtst]

That looks good to me ! Memory measurement look more coherent. Thanks !

[TomTheBear]

Thanks for the response, I will merge the PR into master.