Server Processors

[yah]

hi,

i'm trying different options and i keep getting

Error: For this processor, Event Based Profiling is supported only with Linux kernel version 4.9 or later.

are there any tools i could use to profile the code using centos 7.4?

thanks,

yah

[yah]

HI,

i tried perf using some of the cpu-events

AMDuProfCLI-bin  collect --list cpu-events

listed

  0x00CA     Retired Indirect Branch Instructions Mispredicted

perf stat   -e cycles -e cpu/event=0xca,umask=0x01,inv,cmask=0x01/    ./bots-sparselu  -n  120 -m  501

Performance counter stats for './bots-sparselu -n 120 -m 501':

    58240550701204      cycles                                                     

    58241020036089      cpu/event=0xca,umask=0x01,inv,cmask=0x01/                                  

     439.596918722 seconds time elapsed

  0x00C4     Retired Taken Branch Instructions

perf stat   -e cycles -e cpu/event=0xc4,umask=0x01,inv,cmask=0x01/    ./bots-sparselu  -n  120 -m  501

Performance counter stats for './bots-sparselu -n 120 -m 501':

    58054751953774      cycles                                                     

    39691284251438      cpu/event=0xc4,umask=0x01,inv,cmask=0x01/                                  

     438.316381850 seconds time elapsed

  0x00C1     Retired uops

Performance counter stats for './bots-sparselu -n 120 -m 501':

    58199009753698      cycles                                                     

    39416777313860      cpu/event=0xc1,umask=0x01,inv,cmask=0x01/                                  

     439.075965240 seconds time elapsed

is it ok to use those events with perf? is there an AMD guide i could read to figure out how to use the umask, inv, cmask properties?

thanks,

yah

[yah]

by the way, i'm using centos 7.4 and this is the cpu description

Vendor ID:             AuthenticAMD

CPU family:            23

Model:                 1

Model name:            AMD EPYC 7451 24-Core Processor

Stepping:              2

[swarup]

Hi Yah,

Please refer Family 0x17 OSRR document for the supported events list:

• Open-Source Register Reference for AMD Family 17h Processors (PUB)

You can find the core PMC evnets list in the section "2.1.11.3 Core Performance Monitor Counters".

Each event id is encoded as PMCxABC, where 'ABC' is the event id in hexadecimal, i.e. 0xABC, which you can use with -e as "event=0xABC"

Each event will have umask to choose a one or more hardware event attributes. Each valid bit stands for one attribute whenever applicable.

To choose bit-0, bit-2, you need to set umask=0x5 and so on. If a bit is not applicable, then that bit must be set to 0. For some events, none of the attribute bits are applicable, then set umask=0x0. Check the Bits table for each event to find out exact umask for each each PMC event.

                Swarup

[swarup]

Hi Yah,

You are using CentOS 7.4 on EPYC processor. The (default) kernel version for CentOS 7.4 is 3.10.0, which doesn't support AMD EPYC processors and all AMD family 23 (0x17) processors. Kernel 4.9.x or later supports AMD family 23 processors.

So, if you try to use any family 23 specific events on CentOS7.x, it would not work.

                    Swarup

[yah]

Hi,

i was wondering if you could give tell me what I'm doing wrong. I tried the following

perf stat  -e cpu/event=0x052,umask=0x0 ./bots-sparselu_base.gcc  -n  120 -m  501

and the following error comes up

event syntax error: '..0x052,umask=0x0'

                                  \___ parser error

Run 'perf list' for a list of valid events

Usage: perf stat [<options>] [<command>]

    -e, --event <event>   event selector. use 'perf list' to list available events

I tried -e without cpu and it did not work. We are using umask set to 0 because we want to get both attributes  ( MabMchCnt and DataPipeSwPfDcHit) information. Could you tell us what we are doing wrong?

thanks,

yah

[swarup]

You missed the closing '/' in your command, which caused the parsing error.

Also, I can see you are using 'umask=0x0' for the event 0x52. If you are trying this on a Family 0x17 processor, then as per the OSRR, umask must have at least 1 bit set out of bit0, bit1. So the meaningful umask values are: 0x1, 0x2, 0x3.

Try the following command:

perf stat  -e cpu/event=0x052,umask=0x1/ ./bots-sparselu_base.gcc  -n  120 -m  501

[yah]

thanks for the explanation. it is working now

yah

[yah]

thanks for the tips