MSRs read time resolution

[Davide_Basilio_B_1]

Hello all, I am trying to understand how fast I can get updated information from the IA32_THERM_STATUS msr to monitor the cores temperature. I am using Linux 3.19 and the msr kernel module on a Haswell laptop core i7 processor. It seems to work fine down to a sampling time of 1ms, but if I go below that it seems that the values only change every 1ms anyway. For instance, if I sample at 500us there seem to be always pairs of the same temperature value. I looked at the msr kernel code but I did not find anything that would limit the time resolution there (it pretty much just translates the pread to assembly code). Is there a limitation at the hardware level? Thanks for any hints; I apologize if the information is somewhere and I missed it.

[McCalpinJohn]

Intel documentation mentions the 1 millisecond update time for the power and temperature MSRs in a couple of places.  I have seen it in the chapter on power monitoring in Volume 3 of the SW Developers Manual, and I think it is also in the Uncore Performance Monitoring Guides for the various Xeon E5 processors.

[Davide_Basilio_B_1]

Thanks John for your reply. So probably 1ms is as fast as one can go due to the hardware.

John D. McCalpin wrote:

Intel documentation mentions the 1 millisecond update time for the power and temperature MSRs in a couple of places.  I have seen it in the chapter on power monitoring in Volume 3 of the SW Developers Manual, and I think it is also in the Uncore Performance Monitoring Guides for the various Xeon E5 processors.

[McCalpinJohn]

Yes.   One of the manuals says that these power and temperature MSR values are only updated about once per millisecond, so querying them more frequently is just extra overhead.