config option to specify CPU fingerprinting fallback (dmidecode may be suboptimal)

[MorphBonehunter]

Nomad version

Output from nomad version

Nomad v1.7.1
BuildDate 2023-12-08T18:11:21Z
Revision 608e719430038cdeb5fe108536d90cf88a8540e3

Operating system and Environment details

Archlinux on amd64, virtualized KVM/QEMU

Issue

As suggested from @shoenig in #19412 this is an extra issue for the new dmidecode fallback for CPU fingerprinting starting with 1.7.
The usage could lead to suboptimal MHz readings.

Reproduction steps

In virtual environments, at least KVM with qemu virtual CPU Models, which may be default selected for new VMs in different orchestrators, the dmidecode exposes "wrong" MHz while lscpu / /proc/cpuinfo do expose much better readings.
I have access for example to environments managed with Proxmox, oVirt and Virtuozzo and there these CPUs are use (compatibility stuff).

Maybe the following example is an edge case, but this shows how the dmidecode readings could lead to more Mhz than the host could provide.

In my homelab environment i have an (realy old) Intel(R) Celeron(R) CPU 847 @ 1.10GHz with two cores.
This system use QEMU/KVM to host some VMs with emulated CPU "Intel Xeon E312xx (Sandy Bridge, IBRS update)" on which the MHz detection was ok (2195 MHz) before the upgrade to 1.7.1.
After the upgrade the Frequency was 0 MHz until i install dmidecode as suggested in #19412.

Now I have 2x 2000 MHz because of the dmidecode reporting (only the relevant Information):

Processor Information
        Socket Designation: CPU 0
        Type: Central Processor
        Family: Other
        Manufacturer: QEMU
        Version: pc-q35-4.0
        Max Speed: 2000 MHz
        Current Speed: 2000 MHz
        Core Count: 1
        Thread Count: 1

Processor Information
        Socket Designation: CPU 1
        Type: Central Processor
        Family: Other
        Manufacturer: QEMU
        Version: pc-q35-4.0
        Max Speed: 2000 MHz
        Current Speed: 2000 MHz
        Core Count: 1
        Thread Count: 1

The lscpu -Je=MHZ and grep -i mhz /proc/cpuinfo still shows the real numbers:

{
   "cpus": [
      {
         "mhz": 1097,5060
      },{
         "mhz": 1097,5060
      }
   ]
}

cpu MHz         : 1097.506
cpu MHz         : 1097.506

So in this Case the detected MHz per dmidecode is nearly 2 times what the host real provides.

Expected Result

The Mhz should be detected around 2195.

Actual Result

The MHz is detected as 4000.

[roman-vynar]

A follow up feedback.
Looks like 1.7.2 works fine now, however there is a slight difference in the http call response to read a node attributes per https://developer.hashicorp.com/nomad/api-docs/nodes#read-node

1.3.5 vs 1.7.2

As we can see it reports zeros under Resources and my scripts relied on that.
However, it is still okay under NodeResources so I will use that.

P.S. dmidecode is not needed.

[FibreFoX]

Just my 2 cents: maybe I'm just blind, but demidecode does not give any data on Raspberry Pi (at least Raspbian).

[MorphBonehunter]

After upgrading to 1.7.2 the detection is working again BUT only after uninstalling the dmidecode as this is used before the fallback.
Besides my homelab, this also affects my production systems (also on a hoster which use Virtuozzo/qemu) and produce wrong Mhz as the CPU is a simulated "AMD EPYC Processor (with IBPB)" and dmidecode shows:

Processor Information
        Socket Designation: CPU 0
        Type: Central Processor
        Family: Other
        Manufacturer: Virtuozzo
        Max Speed: 2000 MHz
        Current Speed: 2000 MHz
        Status: Populated, Enabled
        Core Count: 6
        Core Enabled: 6
        Thread Count: 1

So i got for this system an total compute of 12000 MHz with dmidecode installed.
(I can not verify this, but I get the impression that on every simulated CPU the MHz in the DMI is hardcoded to 2000Mhz.)

So the real numbers are:

~ $ lscpu -Je=MHZ
{
   "cpus": [
      {
         "mhz": 2794,7500
      },{
         "mhz": 2794,7500
      },{
         "mhz": 2794,7500
      },{
         "mhz": 2794,7500
      },{
         "mhz": 2794,7500
      },{
         "mhz": 2794,7500
      }
   ]
}
~ $ grep -i mhz /proc/cpuinfo
cpu MHz         : 2794.750
cpu MHz         : 2794.750
cpu MHz         : 2794.750
cpu MHz         : 2794.750
cpu MHz         : 2794.750
cpu MHz         : 2794.750

So total compute is 16764 Mhz which is around 39% more than nomad detects from DMI.

[shoenig]

the dmidecode exposes "wrong" MHz while lscpu / /proc/cpuinfo do expose much better readings.

So the problem we have now is that in some cases /proc/cpuinfo contains better information, while in other cases dmidecode produces better information. For example in EC2 /proc/cpuinfo produces only live frequencies which could be like 800 Mhz at idle or turbo boosted for a brief instant. But from the random sampling of instance types I tried, the dmidecode values set the "Current Speed" to the true base speed which is what we actually want.

I'm not sure what the best thing is to do here. The workaround described above where you choose whether or not to have dmidecode installed isn't ideal; maybe it would be better to have a client configuration value that indicates which fallback method to try first.

I do think the behavior we have now - use dmidecode information if available - is the best default since running in EC2 is very, very common for our users.