On Thu, Dec 4, 2014 at 3:09 AM, Sun, Yongjie <yongjie....@intel.com> wrote: > Dmesg output: > > Performance Events: PEBS fmt1+, SandyBridge events, Intel PMU driver. > PAX: PMU arbitration service v1.0.1 has been started. > sep3_10: PMU collection driver v3.10.3 (EMON INTERNAL) has been loaded. > sep3_10: IDT vector 0x40 will be used for handling PMU interrupts. > clearing PMU state on CPU#15 > clearing PMU state on CPU#8 > > > syst_count output: same like: > > CPU0 G0 8,886,925 8,886,925 unhalted_core_cycles > (scaling 0.00%, ena=1,000,291,924, run=1,000,291,924) > CPU0 G0 0 0 > unhalted_reference_cycles (scaling 0.00%, ena=1,000,291,924, > run=1,000,291,924)
I think both counts are too low. In 1s you got on 9 million cycles..... If the cpu is really busy it should be in the billions. So something is broken. Can you reboot your system AND NOT load the sep driver (VTUNE) and rerun the test? > # 1s ----- > CPU1 G0 2,955,018 2,955,018 unhalted_core_cycles > (scaling 0.00%, ena=1,000,319,397, run=1,000,319,397) > CPU1 G0 0 0 > unhalted_reference_cycles (scaling 0.00%, ena=1,000,319,397, > run=1,000,319,397) > # 1s ----- > CPU2 G0 162,771 162,771 unhalted_core_cycles > (scaling 0.00%, ena=1,000,317,563, run=1,000,317,563) > CPU2 G0 0 0 > unhalted_reference_cycles (scaling 0.00%, ena=1,000,317,563, > run=1,000,317,563) > # 1s ----- > CPU3 G0 318,645 318,645 unhalted_core_cycles > (scaling 0.00%, ena=1,000,314,571, run=1,000,314,571) > CPU3 G0 0 0 > unhalted_reference_cycles (scaling 0.00%, ena=1,000,314,571, > run=1,000,314,571) > # 1s ----- > CPU4 G0 160,486 160,486 unhalted_core_cycles > (scaling 0.00%, ena=1,000,312,981, run=1,000,312,981) > CPU4 G0 0 0 > unhalted_reference_cycles (scaling 0.00%, ena=1,000,312,981, > run=1,000,312,981) > # 1s ----- > CPU5 G0 448,774 448,774 unhalted_core_cycles > (scaling 0.00%, ena=1,000,311,395, run=1,000,311,395) > CPU5 G0 0 0 > unhalted_reference_cycles (scaling 0.00%, ena=1,000,311,395, > run=1,000,311,395) > # 1s ----- > CPU6 G0 149,396 149,396 unhalted_core_cycles > (scaling 0.00%, ena=1,000,307,538, run=1,000,307,538) > CPU6 G0 0 0 > unhalted_reference_cycles (scaling 0.00%, ena=1,000,307,538, > run=1,000,307,538) > # 1s ----- > CPU7 G0 417,315 417,315 unhalted_core_cycles > (scaling 0.00%, ena=1,000,303,946, run=1,000,303,946) > CPU7 G0 0 0 > unhalted_reference_cycles (scaling 0.00%, ena=1,000,303,946, > run=1,000,303,946) > # 1s ----- > CPU8 G0 720,604 720,604 unhalted_core_cycles > (scaling 0.00%, ena=1,000,270,014, run=1,000,270,014) > CPU8 G0 0 0 > unhalted_reference_cycles (scaling 0.00%, ena=1,000,270,014, > run=1,000,270,014) > # 1s ----- > CPU9 G0 1,556,507 1,556,507 unhalted_core_cycles > (scaling 0.00%, ena=1,000,266,053, run=1,000,266,053) > CPU9 G0 0 0 > unhalted_reference_cycles (scaling 0.00%, ena=1,000,266,053, > run=1,000,266,053) > # 1s ----- > CPU10 G0 14,548,713 14,548,713 unhalted_core_cycles > (scaling 0.00%, ena=1,000,266,471, run=1,000,266,471) > CPU10 G0 0 0 > unhalted_reference_cycles (scaling 0.00%, ena=1,000,266,471, > run=1,000,266,471) > # 1s ----- > CPU11 G0 158,015 158,015 unhalted_core_cycles > (scaling 0.00%, ena=1,000,261,804, run=1,000,261,804) > CPU11 G0 0 0 > unhalted_reference_cycles (scaling 0.00%, ena=1,000,261,804, > run=1,000,261,804) > > -----Original Message----- > From: Stephane Eranian [mailto:eran...@googlemail.com] > Sent: Thursday, December 4, 2014 9:48 AM > To: Sun, Yongjie > Cc: perfmon2-devel@lists.sourceforge.net > Subject: Re: [perfmon2] perfmon2 CPU_CLK_UNHALTED:REF_P count error > > hi > > On Thu, Dec 4, 2014 at 2:46 AM, Sun, Yongjie <yongjie....@intel.com> wrote: >> >> I have tried many os like Ubuntu, centos.. >> Kernel versio is: both are the same result 3.5.0-49-generic >> 2.6.32-358.el6.x86_64 >> > Send me the output of: > sudo LIBPFM_VERBOSE=1 ./syst_count -e > unhalted_core_cycles,unhalted_ref_cycles -p -d10 > > Also the output of: dmesg | fgrep PMU > > >> -----Original Message----- >> From: Stephane Eranian [mailto:eran...@googlemail.com] >> Sent: Thursday, December 4, 2014 9:44 AM >> To: Sun, Yongjie >> Cc: perfmon2-devel@lists.sourceforge.net >> Subject: Re: [perfmon2] perfmon2 CPU_CLK_UNHALTED:REF_P count error >> >> Hi, >> >> On Thu, Dec 4, 2014 at 2:39 AM, Sun, Yongjie <yongjie....@intel.com> wrote: >>> Current the issue is: the CPU_CLK_UNHALTED:REF_P couted value is too small >>> than CPU_CLK_UNHALTED:THREAD_P. is it right? From my understanding, it >>> should be almost same. It is none of the workload business. >>> >> Which kernel is this running on? >> UNHALTED_REF_CYCLES can only be measured on a fixed counter. So need kernel >> support for this. >> >> >>> -----Original Message----- >>> From: Stephane Eranian [mailto:eran...@googlemail.com] >>> Sent: Thursday, December 4, 2014 9:35 AM >>> To: Sun, Yongjie >>> Cc: perfmon2-devel@lists.sourceforge.net >>> Subject: Re: [perfmon2] perfmon2 CPU_CLK_UNHALTED:REF_P count error >>> >>> Hi, >>> >>> On Thu, Dec 4, 2014 at 2:31 AM, Sun, Yongjie <yongjie....@intel.com> wrote: >>>> I try use the syst_count: >>>> Result is same >>>> >>>> CPU0 G0 7,188,625 7,188,625 UNHALTED_CORE_CYCLES >>>> (scaling 0.00%, ena=1,999,298,896, run=1,999,298,896) >>>> CPU0 G0 3 3 >>>> UNHALTED_REFERENCE_CYCLES (scaling 0.00%, ena=1,999,336,546, >>>> run=1,999,336,546) >>>> CPU0 G0 7,303,052 7,303,052 >>>> CPU_CLK_UNHALTED:THREAD_P (scaling 0.00%, ena=1,999,348,997, >>>> run=1,999,348,997) >>>> CPU0 G0 545,254 545,254 >>>> CPU_CLK_UNHALTED:REF_P (scaling 0.00%, ena=1,999,364,938, >>>> run=1,999,364,938) >>>> >>> But what is your workload doing and on which CPU does it run. >>> Here it needs to run on CPU0. If you workload blocks and there is nothing >>> else to run on CPU0, the processor goes in halted state, and the events you >>> are measuring stop counting. >>> >>>> -----Original Message----- >>>> From: Stephane Eranian [mailto:eran...@googlemail.com] >>>> Sent: Thursday, December 4, 2014 9:25 AM >>>> To: Sun, Yongjie >>>> Cc: perfmon2-devel@lists.sourceforge.net >>>> Subject: Re: [perfmon2] perfmon2 CPU_CLK_UNHALTED:REF_P count error >>>> >>>> Hi, >>>> >>>> These Python scripts are all not maintained. >>>> I suggest you use the perf_examples/syst_count program instead. >>>> And you need to be root to run per-CPU (system-wide) measurments. >>>> >>>> >>>> On Thu, Dec 4, 2014 at 1:34 AM, Sun, Yongjie <yongjie....@intel.com> wrote: >>>>> Hi, >>>>> I running on Intel SNB E5-2680 >>>>> >>>>> I just used the "libpfm-4.5.0/python/sys.py" >>>>> ./sys.py -e >>>>> UNHALTED_CORE_CYCLES,UNHALTED_REFERENCE_CYCLES,CPU_CLK_UNHALTED:THR >>>>> E >>>>> A >>>>> D >>>>> _P,CPU_CLK_UNHALTED:REF_P -c 0 >>>>> Result: >>>>> CPU0: UNHALTED_CORE_CYCLES 3229292 >>>>> CPU0: UNHALTED_REFERENCE_CYCLES 3 >>>>> CPU0: CPU_CLK_UNHALTED:THREAD_P 3266550 >>>>> CPU0: CPU_CLK_UNHALTED:REF_P 268798 >>>>> CPU0: UNHALTED_CORE_CYCLES 6834444 >>>>> CPU0: UNHALTED_REFERENCE_CYCLES 3 >>>>> CPU0: CPU_CLK_UNHALTED:THREAD_P 6897896 >>>>> CPU0: CPU_CLK_UNHALTED:REF_P 568677 >>>>> CPU0: UNHALTED_CORE_CYCLES 10782920 >>>>> CPU0: UNHALTED_REFERENCE_CYCLES 3 >>>>> CPU0: CPU_CLK_UNHALTED:THREAD_P 10876957 >>>>> CPU0: CPU_CLK_UNHALTED:REF_P 897739 >>>>> -----Original Message----- >>>>> From: Stephane Eranian [mailto:eran...@googlemail.com] >>>>> Sent: Wednesday, December 3, 2014 12:33 AM >>>>> To: Sun, Yongjie >>>>> Cc: perfmon2-devel@lists.sourceforge.net >>>>> Subject: Re: [perfmon2] perfmon2 CPU_CLK_UNHALTED:REF_P count error >>>>> >>>>> Hi, >>>>> >>>>> On Mon, Dec 1, 2014 at 6:24 AM, Sun, Yongjie <yongjie....@intel.com> >>>>> wrote: >>>>>> Hi, all >>>>>> I use the perfmon2 to count events >>>>>> “UNHALTED_CORE_CYCLES,UNHALTED_REFERENCE_CYCLES,CPU_CLK_UNHALTED:THREAD_P,CPU_CLK_UNHALTED:REF_P” >>>>>> and got >>>>> >>>>> which CPU is this running on? >>>>> How long does your test run? >>>>> Are you running in system-wide or per-process mode? >>>>> Need more context to answer your question. >>>>> >>>>>> CPU0: UNHALTED_CORE_CYCLES 5538707 >>>>>> CPU0: UNHALTED_REFERENCE_CYCLES 3 >>>>>> CPU0: CPU_CLK_UNHALTED:THREAD_P 5641229 >>>>>> CPU0: CPU_CLK_UNHALTED:REF_P 419607 >>>>>> CPU0: UNHALTED_CORE_CYCLES 11661876 >>>>>> CPU0: UNHALTED_REFERENCE_CYCLES 3 >>>>>> CPU0: CPU_CLK_UNHALTED:THREAD_P 11813664 >>>>>> CPU0: CPU_CLK_UNHALTED:REF_P 911248 >>>>>> CPU0: UNHALTED_CORE_CYCLES 15682069 >>>>>> CPU0: UNHALTED_REFERENCE_CYCLES 3 >>>>>> CPU0: CPU_CLK_UNHALTED:THREAD_P 15869779 >>>>>> CPU0: CPU_CLK_UNHALTED:REF_P 1246315 >>>>>> >>>>>> Two questions: 1. why UNHALTED_REFERENCE_CYCLES is 3? Too small >>>>>> 2. why CPU0: CPU_CLK_UNHALTED:REF_P is much smaller >>>>>> than CPU_CLK_UNHALTED:THREAD_P? this do not make sence! >>>>>> >>>>>> BTW: I count this for I want to count the CPU Frequency from Intel’s DOC: >>>>>> https://software.intel.com/en-us/articles/measuring-the-average-un >>>>>> h >>>>>> a >>>>>> l >>>>>> t >>>>>> ed-frequency >>>>>> Average frequency = >>>>>> TSC_frequency * (CPU_CLK_UNHALTED.THREAD / CPU_CLK_UNHALTED.REF) >>>>>> >>>>>> So can anybody help on this question? >>>>>> >>>>>> >>>>>> ------------------------------------------------------------------ >>>>>> - >>>>>> - >>>>>> - >>>>>> - >>>>>> -------- Download BIRT iHub F-Type - The Free Enterprise-Grade >>>>>> BIRT Server from Actuate! Instantly Supercharge Your Business >>>>>> Reports and Dashboards with Interactivity, Sharing, Native Excel >>>>>> Exports, App Integration & more Get technology previously reserved >>>>>> for billion-dollar corporations, FREE >>>>>> http://pubads.g.doubleclick.net/gampad/clk?id=157005751&iu=/4140/ostg. >>>>>> clktrk _______________________________________________ >>>>>> perfmon2-devel mailing list >>>>>> perfmon2-devel@lists.sourceforge.net >>>>>> https://lists.sourceforge.net/lists/listinfo/perfmon2-devel >>>>>> ------------------------------------------------------------------------------ Download BIRT iHub F-Type - The Free Enterprise-Grade BIRT Server from Actuate! Instantly Supercharge Your Business Reports and Dashboards with Interactivity, Sharing, Native Excel Exports, App Integration & more Get technology previously reserved for billion-dollar corporations, FREE http://pubads.g.doubleclick.net/gampad/clk?id=164703151&iu=/4140/ostg.clktrk _______________________________________________ perfmon2-devel mailing list perfmon2-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/perfmon2-devel
