[BUG] Events and Counters not initialized with cc-metric-collector

[Joehoch2]

Dear Likwid-Team,

I have built a new version of the cc-metric-collector and tried to update likwid to 5.3.0. When I start the cc-metric-collector with the -debug option I get the following message:

cc-metric-collector[27367]: ERROR - [/root/likwid-5.3.0/src/perfmon.c:perfmon_init:2109] No such file or directory.  
cc-metric-collector[27367]: Failed to initialize event and counter lists for Intel Xeon Broadwell EN/EP/EX processor  
cc-metric-collector[27367]: ERROR 2024/04/23 13:24:57 [LikwidCollector|/root/cc-metric-collector/collectors/likwidMetric.go:781] [failed to initialize library, error -22]

The error does not occur with version 5.2.2 of likwid. So I thought maybe this is the right contact point to get help.

To Reproduce

• LIKWID command and/or API usage Unfortunately, I have not yet understood which commands or API requests are used to query the metrics for cc-metric-collector. When i run a command with likwid-perfctr it actually seems to work: likwid-perfctr -H -g CLOCK likwid-perfctr -e

• LIKWID version and download source (Github, FTP, package manger, ...) 5.3.0 Github als Tarball

• Operating system CentOS 7.9

• Architecture Broadwell EP

I hope this information is enough. If not please let me know

[TomTheBear]

Thanks for the issue but it seems to be not in the proper location (repository). Since LIKWID seems to work (likwid-perfctr -e works), this problem is probably on the cc-metric-collector side.

The case is odd because this happens when multiple LIKWID versions are installed and the wrong one is picked at runtime but the output clearly states 5.3.0 (/root/likwid-5.3.0/src/perfmon.c:perfmon_init:2109). But for not supporting BroadwellEP, the LIKWID library has to be really ancient.

[Joehoch2]

I dug deeper into this issue. At first i remembered that when you build the cc-metric-collector it downloads a older likwid-version and copies its header files. So i changed that in the makefile to download the latest version, but with no success. After that I compared src/perfmon.c of likwids version 5.3.0 and 5.2.2 and found the codeblock of the error:

    ret = perfmon_init_maps();
    if (ret < 0)    
     {               
         ERROR_PRINT(Failed to initialize event and counter lists for %s, cpuid_info.name);
         HPMfinalize();
         return ret;
     }

I changed the whole codeblock into perfmon_init_maps(); and it works like in version 5.2.2, but i think it is not the purpose, because the return code should be 0 if the function suceeded, I guess. I hope i get that right.

Nevertheless thank you for the response. Shall I copy that issue the other repository?