CPU usage on hyperthreaded Xeon processors

[John Poff]

Hi,

I'm having a discussion with a mainframe guy who is doing some testing on one of my Linux boxes. I have a box with two Xeon 2.8 Ghz CPUs in it, with hyperthreading turned on, and I'm running RedHat Advanced Server 2.1. The box shows four CPUs in all the usual utilities [sar, top, etc.] as a result of the hyperthreading.

My mainframe friend is running a test script that reads a large file. He notes that his script consumes over 99% of one CPU, while the other three show no usage. His contention is that since he is getting half of one physical 2.8 Ghz CPU, his script is running on a virtual 1.4 Ghz CPU. I argue back with him that if nothing else is running on the second virtual CPU, he is getting all of the physical CPU. I'm not a wizard on hyperthreading, but it seems that his CPU usage would always be at 2.4 Ghz. I understand what he is trying to say about getting half the physical CPU, but I don't think that saying he has a 1.4 Ghz CPU is a fair statement. I think he gets a 2.8 Ghz CPU for half as much time.

Opinions? Comments?

JP

[Vitaly Karasik_1]

of course, there is no "1.4 GHz virtual CPU"

related links:

http://arstechnica.com/paedia/h/hyperthreading/hyperthreading-1.html

http://developer.intel.com/software/products/browse/itanium.htm

http://www-106.ibm.com/developerworks/linux/library/l-htl/?ca=dgr-lnxw01

Rgds

Vitaly

[Umapathy S]

JP,

I think the way OS behaves for hypethreading is still limited. The advantage is more felt only when many application run simultaneously.

More here
http://www.2cpu.com/Hardware/ht_analysis/

Some statistics here
http://www.pcworld.com/news/article/0,aid,107492,00.asp

cheers
Umapathy

[Steven E. Protter]

I would think the results would be different if the box was booted off one of the SMP kernels.

For what its worth I think you're instincts are right. I love those boxes you are running, fast and furious.

I'm running the old web hosting business of Red Hat 7.3's stock smp kernel even though I have but one CPU right now, that is the way I'd run things in a multi-cpu environment.

What is your Linux Distribution?

SEP

[John Poff]

Vitaly and Umapathy,

Thanks for the links. That is some good info.

Steven,

I'm running Advanced Server 2.1, using the e25 enterprise kernel,
since I have two CPUs and 8 Gb of RAM. I just built another one
this week, same configuration, except that the CPUs are 3.066 Ghz.
Those are some *FAST* boxes, and I'm having a blast working with them.


My point with my mainframe guy was that if he is running something
on one CPU at 99%, and the other CPUs are idle, then he must be getting
pretty much all of a 2.8 Ghz CPU, and not a 1.4 Ghz virtual CPU. Those
mainframers are used to dividing up CPUs on their system, and maybe it
works that way on their end, but I'm debating the point with this guy and
I just wanted some feedback to see if I was understanding the hyperthreading
correctly.

JP

[Stuart Browne · Accepted solution]

No, Hyperthreading is not like mainframe dividing up clock cycles.

It literally is multiple-executes-per-cycle.

Using one single threaded application won't show one whit of performance boost. Use 10 of them at a time, and you might see some improovement. Using a well written multi-threaded application, you'd see a reasonable amount of improovement.

You might find this useful:

http://www-106.ibm.com/developerworks/linux/library/l-htl/

[Khalid A. Al-Tayaran]

Hi,


I don't think that his statement is right. You are getting 2.8GHz not 1.4GHz. But it depends on the script and the OS....

More info on CPUs here:
http://www.geek.com/procspec/procspec.htm