Open milianw opened 7 years ago
Yes could be due to the HT exclusion workaround. Also try something that runs longer.
You can disable HT with
cputop.py 'thread == 1' offline | sh
(or thread == 0, if cpu 0 is included)
What do I need to get e.g. /sys/devices/system/cpu/cpu0/online
? I don't have the online
pseudo-file.
It's a CONFIG option of the kernel.
But if thread == 1 includes cpu 0, just off line thread == 0 instead. It just has to be either all 0 or all 1 to disable hyper threading. Alternatively you can also disable it in the BIOS.
Hello! I encounter the same problem. I still have the same "division by zero" errors even though I disabled hyper-threading using cputop.py. Are there any other potential problems that could cause this?
I can see all the main metrics when using "perf stat -d my_workload".
Thanks
On my machine, the total clock count is not measured. At the bottom of the toplev output, I also see that I get
11 division by zero errors: IPC CoreIPC CPI ...
, which seems to indicate that the instruction / cycle count fails on my machine? This is odd, considering that cycles/instructionsI then disabled the NMI watchdog and try to launch toplev:
If I just try to run parts of the perf command, I also see the broken output:
If I remove any two of the events, the command starts to produce proper output:
Is this maybe due to hyper threading?