PAPI Forum

Hi,
am a computer science student from germany and i do my internship at the moment.

My job is to do performance analysis of virtualized multicore machines with KVM and compare the results of the benchmarks with the real machine. Another student has already done this with XEN.

Kernel 3.3 introduced a vPMU to KVM and perf-events. Will you support the vPMU in the future or does it already work by using libpfm?

Regards,
Andreas

PAPI doesn't need any extra vPMU support. If the virtualization is done properly PAPI will behave just as it does on real hardware.

We've tested PAPI inside of KVM with a Linux 3.3 guest and a current git-snapshot of Qemu. It works for getting aggregate counts, but the kernel still does not implement overflow or sampling.

Thank you very much for your answer.

I am using the latest Alpha of Fedora 17. I already got the information that most of the code related to the vPMU is in the perf-event subsystem.

When there is no need to adapt PAPI to the vPMU, than this adaption can only reside in on of PAPI's backends like perfctr or perf-event?

I already tried oprofile with Kernel 3.3.1 with qemu 1.0 and it wan't working. I now try the oprofile perf-event branch and than PAPI.

I hope i don't bother you with my questions, but is the GIT qemu really needed for the vPMU or is the latest 3.3.1 Kernel sufficient?

I experimented a lot with Fedora. I tried to install the most actual Distribution, wich is Fedora 17 Beta, togehter with the kernel and Qemu from Rawhide and PAPI via git. Now i understand the problem. Papi_avail detects the events but when i run the examples i nearly always get error -8.

You have written about the problems with overflow. Could it be that the NMI generated by the overflow isn't injected into the VM? If so, the problem would be on the KVM(Kernel) or QEMU side.