libpcp internal error when recording metrics with pmrep

[myllynen]

These three commands work:

pmrep -i 1 -o archive -F ./foo proc.memory.vmhwm,,,
pmrep -i 1 -o archive -F ./foo proc.memory.vmhwm,,,MB
pmrep -i 1,2 -o archive -F ./foo proc.memory.vmhwm,,,

But this fails with unexpected error message:

pmrep -i 1,2 -o archive -F ./foo proc.memory.vmhwm,,,MB
Error: _pmi_stuff_value: vset realloc:: malloc(-197760) failed: Cannot allocate memory

I'm not sure is pmrep(1) doing the right thing here but I suspect libpcp error handling could be improved? After that we could see if changes are warranted on the pmrep side as well.

Thanks.

[natoscott]

Hi Marko,

I think there's two problems here - one in the C library (libpcp_import) and the other in the python wrapper. I'll look into the first but I wonder if you could have a poke around the python library code for the latter in parallel?

The C library problem (and cause of the diagnostic message) appears to be a failure to handle an error code (PM_ERR_CONV) correctly when constructing the pmResult structure being written into the archive.

The python problem is that we're passing a floating point number like "16.78515625" (as a string) for a type u32 metric via the pmiPutValue(3) routine. I expect this is because of the change in scale that the additional "MB" part of your pmrep invocation induces. This may have been a by-product of the changes in integer / floating point handling between python 2 and 3 but that's a wild guess.

$ pminfo 3.24.35 -d

proc.memory.vmhwm
    Data Type: 32-bit unsigned int  InDom: 3.9 0xc00009
    Semantics: instant  Units: Kbyte

[myllynen]

Thanks for taking care of the libpcp part, I've fixed pmrep/pmconfig.py with:

3806df8