Re: Using nr_cpus breaks no_turbo setting of pstate driver

From: Bhatnagar, Rishabh
Date: Thu Dec 01 2022 - 13:32:25 EST



On 12/1/22 8:56 AM, srinivas pandruvada wrote:

*CAUTION*: This email originated from outside of the organization. Do not click links or open attachments unless you can confirm the sender and know the content is safe.



On Wed, 2022-11-30 at 17:03 -0800, Bhatnagar, Rishabh wrote:

On 11/30/22 8:40 AM, srinivas pandruvada wrote:
CAUTION: This email originated from outside of the organization. Do not click links or open attachments unless you can confirm the sender and know the content is safe.



Hi Rishabh,

I think this is because of request of SMT and the the following CPUID
flag is set to false.
# cpuid -1 -l 6 | grep ignoring
ignoring idle logical processor HWP req = false

No this flag is not there in my cpuid output. here are the HWP flags i
could see:

       HWP base registers                      = true
       HWP notification                        = false
       HWP activity window                     = true
       HWP energy performance preference       = true
       HWP package level request               = true

The way CPUs are numbered, the SMT starts at number 24. So, nr_cpus=24
didn't bring up any SMT for driver to control.

I tried experiment with nr_cpus=25. Seems like the core which has both
cpus enabled follows the turbo limits and rest do not.
Can you explain how SMT is related to cpus staying in turbo frequency?
The frequency controls happens at core level not at SMT level. The default  max powerup SMT HWP max request is 0xFF.
Since this is not controlled by the intel_pstate, the request from SMT (which was not part of powerup CPUs) is higher.
The higher one wins.

Thanks,
Srinivas

So what i understand is that even though SMT is not enabled because of reduced cpus it has a initial default vote of 255. (Does this vote show up in any register in turbostat output?)

This default request always overpowers what we set in HWP_REQUEST.

Isn't this an issue with the implementation since if we boot with reduced cpus they don't come down from turbo frequencies unless msr is touched? Or this behavior desired?

Also one more observation i couldn't explain was if i manually set smt/control off dynamically (nr_cpus=24) then also the cpus come down from turbo frequencies. How does disabling smt help here?


Also if i dynamically offline 24 cpus then i don't see this issue. It
only appears when booted with 24 cpus.

^CPackage       Core    CPU     Avg_MHz Busy%   Bzy_MHz TSC_MHz IRQ
SMI     POLL    C1      POLL%   C1%     CPU%c1  CPU%c6 CoreTmp PkgTmp
Pkg%pc2 Pkg%pc6 PkgWatt RAMWatt PKG_%   RAM_%
-       -       -       4379    99.64   4395    3800 3780 0
0       0       0.00    0.00    0.36    0.00    88 88      0.00
0.00    479.85  66.88   0.00    0.00
*0       0       0       3788    99.68   3800    3800 146 0
0       0       0.00    0.00    0.32    0.00    72 88      0.00
0.00    239.97  30.38   0.00    0.00**
**0       0       24      3788    99.68   3800    3800 146     0
0       0       0.00    0.00    0.32*
0       1       1       4372    99.63   4388    3800 146 0
0       0       0.00    0.00    0.37    0.00    84
0       2       2       4372    99.63   4388    3800 146 0
0       0       0.00    0.00    0.37    0.00    88
0       4       3       4372    99.63   4388    3800 174 0
0       0       0.00    0.00    0.37    0.00    83
0       9       4       4372    99.63   4388    3800 146 0
0       0       0.00    0.00    0.37    0.00    86
0       10      5       4372    99.63   4388    3800 150 0
0       0       0.00    0.00    0.37    0.00    87
0       11      6       4372    99.63   4388    3800 146 0
0       0       0.00    0.00    0.37    0.00    87
0       16      7       4372    99.63   4388    3800 146 0
0       0       0.00    0.00    0.37    0.00    84
0       17      8       4372    99.64   4388    3800 146 0
0       0       0.00    0.00    0.36    0.00    86
0       18      9       4372    99.64   4388    3800 146 0
0       0       0.00    0.00    0.36    0.00    84
0       19      10      4372    99.63   4388    3800 146 0
0       0       0.00    0.00    0.37    0.00    84
0       27      11      4372    99.62   4388    3800 148 0
0       0       0.00    0.00    0.38    0.00    84
1       1       12      4484    99.64   4500    3800 147 0
0       0       0.00    0.00    0.36    0.00    69 73      0.00
0.00    239.83  36.49   0.00    0.00
1       2       13      4484    99.64   4500    3800 147 0
0       0       0.00    0.00    0.36    0.00    73
1       4       14      4484    99.64   4500    3800 146 0
0       0       0.00    0.00    0.36    0.00    67
1       8       15      4484    99.65   4500    3800 146 0
0       0       0.00    0.00    0.35    0.00    68
1       9       16      4484    99.64   4500    3800 146 0
0       0       0.00    0.00    0.36    0.00    70
1       10      17      4484    99.64   4500    3800 146 0
0       0       0.00    0.00    0.36    0.00    68
1       11      18      4484    99.64   4500    3800 146 0
0       0       0.00    0.00    0.36    0.00    72
1       16      19      4484    99.65   4500    3800 146 0
0       0       0.00    0.00    0.35    0.00    69
1       18      20      4484    99.65   4500    3800 146 0
0       0       0.00    0.00    0.35    0.00    64
1       19      21      4484    99.65   4500    3800 146 0
0       0       0.00    0.00    0.35    0.00    67
1       24      22      4484    99.64   4500    3800 146 0
0       0       0.00    0.00    0.36    0.00    72
1       27      23      4484    99.64   4500    3800 240 0
0       0       0.00    0.00    0.36    0.00    65

The default power up MSR_HWP_REQ.max is 0xFF, this will also play role
in resolving the max frequency. Here driver is disabling turbo by
controlling MSR_HWP_REQ.max.

If you want to disable turbo here with reduced nr_cpus, you can set bit
38 of MSR 0x1a0 using msr tools.
Trying to get more understanding on why does the default way of setting
no_turbo not work with reduced cpus?
I verified this works for nr_cpus=24/25.

Thanks,
Srinivas


On Wed, 2022-11-30 at 08:10 -0800, Bhatnagar, Rishabh wrote:
"Having some issue with my mails failing to deliver. So trying once
again after disabling HTML."


Hi Srinivas/Len,


I have been doing some testing on m5zn.metal AWS instances which
provides 48 vcpus and 192GB memory.

Test includes loading all cpus with maximum capacity (using stress)
and
observing the frequency after enabling and disabling turbo
frequencies.
Intel_pstate is enabled with hwp.


lscpu output for reference:

*# lscpu*
Architecture:        x86_64
CPU op-mode(s):      32-bit, 64-bit
Byte Order:          Little Endian
CPU(s):              48
On-line CPU(s) list: 0-47
Thread(s) per core:  2
Core(s) per socket:  12
Socket(s):           2
NUMA node(s):        2
Vendor ID:           GenuineIntel
CPU family:          6
Model:               85
Model name:          Intel(R) Xeon(R) Platinum 8252C CPU @ 3.80GHz
Stepping:            7
CPU MHz:             3800.000
CPU max MHz:         4500.0000
CPU min MHz:         1200.0000
BogoMIPS:            7600.00
Virtualization:      VT-x
L1d cache:           32K
L1i cache:           32K
L2 cache:            1024K
L3 cache:            25344K
NUMA node0 CPU(s):   0-11,24-35
NUMA node1 CPU(s):   12-23,36-47
Flags:               fpu vme de pse tsc msr pae mce cx8 apic sep mtrr
pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm
pbe
syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts
rep_good nopl xtopology nonstop_tsc cpuid aperfmperf pni pclmulqdq
dtes64 monitor ds_cpl vmx smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm
pcid
dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave
avx
f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb cat_l3 cdp_l3
invpcid_single intel_ppin ssbd mba ibrs ibpb stibp ibrs_enhanced
tpr_shadow vnmi flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1
avx2 smep bmi2 erms invpcid cqm mpx rdt_a avx512f avx512dq rdseed adx
smap clflushopt clwb intel_pt avx512cd avx512bw avx512vl xsaveopt
xsavec
xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local
dtherm
ida arat pln pts hwp hwp_act_window hwp_epp hwp_pkg_req pku ospke
avx512_vnni md_clear flush_l1d arch_capabilities

When testing with all cpus online I don't see any issue. CPUs reach
turbo frequency when intel_pstate/no_turbo is disabled and stay below
turbo frequencies when no_turbo is enabled as expected.

However, issue occurs when I boot with nr_cpus argument specifying
less
than 48 cpus. e.g. cmdline:


*/# cat /proc/cmdline/*/
/mmio_stale_data=off mds=off tsx=off tsx_async_abort=off
intel_idle.max_cstate=0
BOOT_IMAGE=/boot/vmlinuz-5.10.154-119.671.amzn2.x86_64
root=UUID=d8605abb-d6cd-4a46-a657-b6bd206da2ab ro console=tty0
console=ttyS0,115200n8 net.ifnames=0 biosdevname=0
nvme_core.io_timeout=4294967295 rd.emergency=poweroff
rd.shell=0*nr_cpus=24 intel_idle.max_cstate=1 processor.max_cstate=1*

# echo 1 > /sys/devices/system/cpu/intel_pstate/no_turbo
# turbostat


^CPackage       Core CPU     Avg_MHz Busy%   Bzy_MHz TSC_MHz IRQ
SMI     POLL C1      POLL%   C1%     CPU%c1  CPU%c6 CoreTmp PkgTmp
Pkg%pc2 Pkg%pc6 PkgWatt RAMWatt PKG_%   RAM_%
-       -       -       4370    99.76   4381 3800    6374 0
0       0       0.00    0.00    0.24    0.00    96 96      0.00
0.00    479.79  69.20   0.00    0.00
0       0       0       4292    99.93   4295 3800    266 0
0       0       0.00    0.00    0.07    0.00    88 96      0.00
0.00    239.89  33.62   0.00    0.00
0       2       1       4292    99.92   4295 3800    270 0
0       0       0.00    0.00    0.08    0.00    92
0       4       2       4292    99.92   4295 3800    264 0
0       0       0.00    0.00    0.08    0.00    92
0       8       3       4292    99.93   4295 3800    268 0
0       0       0.00    0.00    0.07    0.00    90
0       9       4       4292    99.93   4295 3800    264 0
0       0       0.00    0.00    0.07    0.00    91
0       10      5       4292    99.92   4295 3800    264 0
0       0       0.00    0.00    0.08    0.00    92
0       11      6       4292    99.92   4295 3800    265 0
0       0       0.00    0.00    0.08    0.00    96
0       17      7       4292    99.93   4295 3800    266 0
0       0       0.00    0.00    0.07    0.00    91
0       19      8       4292    99.93   4295 3800    266 0
0       0       0.00    0.00    0.07    0.00    91
0       25      9       4292    99.92   4295 3800    264 0
0       0       0.00    0.00    0.08    0.00    91
0       26      10      4292    99.92   4295 3800    264 0
0       0       0.00    0.00    0.08    0.00    93
0       27      11      4292    99.92   4295 3800    264 0
0       0       0.00    0.00    0.08    0.00    94
1       1       12      4449    99.61   4466 3800    264 0
0       0       0.00    0.00    0.39    0.00    80 88      0.00
0.00    239.90  35.58   0.00    0.00
1       2       13      4448    99.60   4466 3800    264 0
0       0       0.00    0.00    0.40    0.00    86
1       3       14      4449    99.61   4466 3800    267 0
0       0       0.00    0.00    0.39    0.00    81
1       8       15      4449    99.61   4466 3800    264 0
0       0       0.00    0.00    0.39    0.00    79
1       9       16      4448    99.60   4466 3800    264 0
0       0       0.00    0.00    0.40    0.00    79
1       10      17      4449    99.61   4466 3800    264 0
0       0       0.00    0.00    0.39    0.00    83
1       17      18      4448    99.60   4466 3800    264 0
0       0       0.00    0.00    0.40    0.00    82
1       18      19      4449    99.61   4466 3800    281 0
0       0       0.00    0.00    0.39    0.00    80
1       19      20      4449    99.61   4466 3800    264 0
0       0       0.00    0.00    0.39    0.00    83
1       20      21      4448    99.59   4466 3800    264 0
0       0       0.00    0.00    0.41    0.00    83
1       25      22      4448    99.60   4466 3800    264 0
0       0       0.00    0.00    0.40    0.00    81
1       27      23      4448    99.59   4466 3800    265 0
0       0       0.00    0.00    0.41    0.00    88

/
/

As part of the turbostat output I could see that pstate driver is
setting the hints correctly in MSR_HWP_REQUEST. But somehow these
limits
are not being honored by HWP logic.


cpu0: MSR_PM_ENABLE: 0x00000001 (HWP)
cpu0: MSR_HWP_CAPABILITIES: 0x070c262d (high 45 guar 38 eff 12 low 7)
*cpu0: MSR_HWP_REQUEST: 0x00002626 (min 38 max 38 des 0 epp 0x0
window
0x0 pkg 0x0)*
cpu0: MSR_HWP_REQUEST_PKG: 0x8000ff00 (min 0 max 255 des 0 epp 0x80
window 0x0)
cpu0: MSR_HWP_STATUS: 0x00000004 (No-Guaranteed_Perf_Change,
No-Excursion_Min)

I tried the same experiment after disabling hwp by
specifying*intel_pstate=no_hwp*in cmdline. In this case the pstate
driver seems to be selecting pstates correctly.

^CPackage       Core CPU     Avg_MHz Busy%   Bzy_MHz TSC_MHz IRQ
SMI     POLL C1      POLL%   C1%     CPU%c1  CPU%c6 CoreTmp PkgTmp
Pkg%pc2 Pkg%pc6 PkgWatt RAMWatt PKG_%   RAM_%
-       -       -       3800    100.00  3800 3800    1646 0
0       0       0.00    0.00    0.00    0.00    85 85      0.00
0.00    328.38  62.14   0.00    0.00
0       0       0       3800    100.00  3800 3800    66 0
0       0       0.00    0.00    0.00    0.00    85 85      0.00
0.00    175.42  36.96   0.00    0.00
0       0       24      3800    100.00  3800 3800    70 0
0       0       0.00    0.00    0.00
0       1       1       3800    100.00  3800 3800    65 0
0       0       0.00    0.00    0.00    0.00    82
0       2       2       3800    100.00  3800 3800    67 0
0       0       0.00    0.00    0.00    0.00    84
0       3       3       3800    100.00  3800 3800    67 0
0       0       0.00    0.00    0.00    0.00    83
0       4       4       3800    100.00  3800 3800    68 0
0       0       0.00    0.00    0.00    0.00    85
0       8       5       3800    100.00  3800 3800    68 0
0       0       0.00    0.00    0.00    0.00    83
0       9       6       3800    100.00  3800 3800    65 0
0       0       0.00    0.00    0.00    0.00    85
0       10      7       3800    100.00  3800 3800    65 0
0       0       0.00    0.00    0.00    0.00    82
0       11      8       3800    100.00  3800 3800    65 0
0       0       0.00    0.00    0.00    0.00    84
0       17      9       3800    100.00  3800 3800    65 0
0       0       0.00    0.00    0.00    0.00    83
0       18      10      3800    100.00  3800 3800    65 0
0       0       0.00    0.00    0.00    0.00    83
0       27      11      3800    100.00  3800 3800    65 0
0       0       0.00    0.00    0.00    0.00    85
1       0       12      3800    100.00  3800 3800    65 0
0       0       0.00    0.00    0.00    0.00    66 68      0.00
0.00    152.95  25.18   0.00    0.00
1       2       13      3800    100.00  3800 3800    65 0
0       0       0.00    0.00    0.00    0.00    65
1       4       14      3800    100.00  3800 3800    65 0
0       0       0.00    0.00    0.00    0.00    68
1       8       15      3800    100.00  3800 3800    65 0
0       0       0.00    0.00    0.00    0.00    63
1       9       16      3800    100.00  3800 3800    65 0
0       0       0.00    0.00    0.00    0.00    65
1       10      17      3800    100.00  3800 3800    65 0
0       0       0.00    0.00    0.00    0.00    64
1       17      18      3800    100.00  3800 3800    66 0
0       0       0.00    0.00    0.00    0.00    63
1       18      19      3800    100.00  3800 3800    66 0
0       0       0.00    0.00    0.00    0.00    63
1       19      20      3800    100.00  3800 3800    65 0
0       0       0.00    0.00    0.00    0.00    63
1       20      21      3800    100.00  3800 3800    66 0
0       0       0.00    0.00    0.00    0.00    66
1       25      22      3800    100.00  3800 3800    65 0
0       0       0.00    0.00    0.00    0.00    64
1       27      23      3800    100.00  3800 3800    67 0
0       0       0.00    0.00    0.00    0.00    67

[root@ip-172-31-36-94 ec2-user]# cat
/sys/devices/system/cpu/intel_pstate/no_turbo
1


Thanks,
Rishabh