Intel Xeon P-State vs. CPUFreq Scaling Tests Show Some Odd Numbers

[phoronix]

Phoronix: Intel Xeon P-State vs. CPUFreq Scaling Tests Show Some Odd Numbers

As we've seen a lot of variation in results with different Intel processors when switching between the Intel P-State and CPUFreq scaling drivers and the different governors, here's some tests when using a 16 thread (eight core + HT) Haswell-EP Xeon processor and testing the different CPU frequency scaling settings in Fedora 21...

http://www.phoronix.com/vr.php?view=MTgzNzY

[johnc]

Maybe Intel will actually get these things working some day.

But with them only making about 500 trillion dollars in pure profit every quarter, I can see how they just have no money in the budget to work on this.

I wonder how frequency scaling governors perform in Windows? Probably flawlessly, no doubt.

[Veske]

It is weird though. Why would they not get it right if most of the servers which require a lot of power from CPU also are on GNU/Linux machines...

[johnc]

It is weird though. Why would they not get it right if most of the servers which require a lot of power from CPU also are on GNU/Linux machines...

Because they have a monopoly?

Though I wonder if machines used in those functions even have frequency scaling enabled?

[curaga]

I wonder how frequency scaling governors perform in Windows? Probably flawlessly, no doubt.

Remember the news how Win8 was required to get proper performance from Bulldozer? That does not inspire confidence in the Windows cpu governor.

[ArchLinux]

Maybe Intel will actually get these things working some day.

But with them only making about 500 trillion dollars in pure profit every quarter, I can see how they just have no money in the budget to work on this.

Exactly! They get nothing right and this is the reason they make only 500 trillion dollars... wait, what?

[gilboa]

Maybe Intel will actually get these things working some day.

But with them only making about 500 trillion dollars in pure profit every quarter, I can see how they just have no money in the budget to work on this.

I wonder how frequency scaling governors perform in Windows? Probably flawlessly, no doubt.

You obviously have no idea what you're talking about.

Hint: Scaling governor that works best across different - let alone contradicting workloads (E.g. DB vs. DPI vs. VM workloads), is nearly impossible.
Add cooling limitations, TDP, BIOS settings [1], throttling and you have a job from hell.

Oh, and from experience, Windows *server* scaling governor on Windows requires a lot of hand tuning by the server manufacturer (E.g. HP, Dell) to get decent performance. And trust me, having a 50K$ server overheat and start throttling is *not* amusing.

But hey, you know best, right?

- Gilboa
* If you have seen the CPU/Cooling configuration menu in a recent server or workstation you'll know what I mean.

[toyotabedzrock]

There must be a thread timing issue in GCC that is causing that result. It is so dramatic they really need to look into this as it could be a huge performance gain.

[ArchLinux]

You obviously have no idea what you're talking about.

Hint: Scaling governor that works best across different - let alone contradicting workloads (E.g. DB vs. DPI vs. VM workloads), is nearly impossible.
Add cooling limitations, TDP, BIOS settings [1], throttling and you have a job from hell.

Oh, and from experience, Windows *server* scaling governor on Windows requires a lot of hand tuning by the server manufacturer (E.g. HP, Dell) to get decent performance. And trust me, having a 50K$ server overheat and start throttling is *not* amusing.

But hey, you know best, right?

- Gilboa
* If you have seen the CPU/Cooling configuration menu in a recent server or workstation you'll know what I mean.

+1 +1.