Known Issues

Known problems

Known issues are always related to the latest stable release. Please have also a look at the issue tracker about limitations and bugs.

The Intel Nehalem / Westmere Flop events are executed events. This causes that on all Nehalem processors the flop events overcount by about 5%
The Intel SandyBridge / IvyBridge Flop events are highly overcounting as soon as the required data is not in L1 cache. Do not trust them.

Reading of the RAPL counters can overflow. This was already fixed in the master branch.

The likwid-agent application in 4.0.0 is experimental. It assumes that gmetric has the --group option (version 3.2.0 or later). Socket-wide counters are read only for the first socket. Spaces in metric names. If you want to use it, take src/applications/likwid-agent.lua from the master branch and copy it to your source tree before installing.
If you use a system-wide installation of hwloc in your application, there will be problems when using likwid because it ships with a bleeding edge version of hwloc. If you link at first with system hwloc and then with LIKWID, LIKWID cannot get the CPU model number and probably other important info. If you link first with LIKWID and then with the system hwloc, you cannot access hwloc structures anymore, because the hwloc header included does not fit the used hwloc.
Don't use pinning for likwid-perfctr if your application pins itself. One example for this is likwid-bench. In order to measure likwid-bench, run likwid-bench without likwid-perfctr and check which CPU cores are used to set a valid cpu list for likwid-perfctr -c <list> -g <events> -m likwid-bench ...
For the Fortran90 interface, there could be symbol visibility problems. I added a patch to fix that on the release page (https://github.com/RRZE-HPC/likwid/releases/download/likwid-4.0.1/likwid-4.0.1-f90-visibility.patch)