Differences in host average between 4 and 8 processor computers

Message boards : Number crunching : Differences in host average between 4 and 8 processor computers

To post messages, you must log in.

AuthorMessage
Ed Machak

Send message
Joined: 10 Nov 16
Posts: 7
Credit: 17,339,411
RAC: 0
Message 96637 - Posted: 19 May 2020, 22:53:59 UTC

Hello,

I have four similar four core CPU computers running Rosetta. They returning similar CPU benchmarks, 15,000 floating point and 39,000 integer computations per CPU. Predicted runtimes are in the 7 to 9 hour range for each of the four processors.

A 5th computer, an 8 CPU Lenovo laptop, has benchmarks of 12,000 floating point and 69,000 integer computations per CPU. Predicted runtimes are also in the 7 to 9 hour range for each of the 8 processors.

The typical host averages of the 4-CPU computers is 3400 but the laptop is only 1800. With double the number of processors why is the 8 CPU host average only ½ of the four CPU computers?

Thank you,

Ed
ID: 96637 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Tomcat雄猫

Send message
Joined: 20 Dec 14
Posts: 180
Credit: 5,386,173
RAC: 0
Message 96638 - Posted: 20 May 2020, 0:01:02 UTC - in response to Message 96637.  
Last modified: 20 May 2020, 0:08:43 UTC

Hello,

I have four similar four core CPU computers running Rosetta. They returning similar CPU benchmarks, 15,000 floating point and 39,000 integer computations per CPU. Predicted runtimes are in the 7 to 9 hour range for each of the four processors.

A 5th computer, an 8 CPU Lenovo laptop, has benchmarks of 12,000 floating point and 69,000 integer computations per CPU. Predicted runtimes are also in the 7 to 9 hour range for each of the 8 processors.

The typical host averages of the 4-CPU computers is 3400 but the laptop is only 1800. With double the number of processors why is the 8 CPU host average only ½ of the four CPU computers?

Thank you,

Ed


A couple of things here.
Firstly, predicted run-times, actual run-times, and estimated computation size don't mean anything in Rosetta. Rosetta at home is different from other projects in that instead of running a task to completion, Rosetta runs a task for a set amount of time. Faster computers get more work done in the amount of time, and get a higher RAC.
Secondly, Rosetta@home relies on floating point performance, not integer performance.
Now that we've addressed those two points let's look at the issue at hand. The reason your 8 core machine is getting less credits is because it is doing less work (running slower). looking at similar tests, like the Junior_HalfRoid_design6_cart_COVID-19_SAVE_ALL_OUT_IGNORE_THE_REST tasks (Rosetta tasks are very varied, so this is not a good comparison)
Your 4 core CPU:
Junior_HalfRoid_design6_cart_COVID-19_SAVE_ALL_OUT_IGNORE_THE_REST_0pm3zs3b_929581_19_0
https://boinc.bakerlab.org/rosetta/result.php?resultid=1184183983
You got 276.71 credits in 29720.2 seconds
Looking at the stderr output, your machine completed 5 "decoys"

Your 8 core machine (note, your "8 CPU" machine does NOT have 8 physical cores, it has 4, like your 4 core machines. It has 8 threads, but only 4 cores. This is called SMT, or hyper-threading in Intel-speak. HT/SMT allows you to have more than one thread on a physical core. In most cases, HT gives a boost of up to 30% in multi-core performance):
Junior_HalfRoid_design6_cart_COVID-19_SAVE_ALL_OUT_IGNORE_THE_REST_6ko7ze9y_929388_20_0
https://boinc.bakerlab.org/rosetta/result.php?resultid=1184688076
You got 59.45 credits in 22037.7 seconds
Looking at the stderr output, your machine completed 1 "decoys"

Looking at other similar tasks we see this is rather consistent. Your 8 core machine is actually running at around 1/4 the speed of your 4 core machine per core.
I think your 8 thread machine may be thermal throttling.
Why don't the benchmarks agree?
I, for one, I find the benchmarks to be rather inaccurate. Modern CPUs boost (run faster than stock) if the thermals are favourable. The benchmark only stresses one core of your CPU and doesn't generate a lot of heat. Hence, the benchmarks show you the absolute best-case scenario for the tested CPU (Max boosts). Actual speeds will very often be slower than the benchmarks (again, I have to point out Rosetta tasks rely on the floating point performance). Your laptop may be running hot enough for the CPU to thermal throttle, which makes it perform worse than stock.
ID: 96638 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Profile Grant (SSSF)

Send message
Joined: 28 Mar 20
Posts: 1681
Credit: 17,854,150
RAC: 18,215
Message 96650 - Posted: 20 May 2020, 9:13:56 UTC - in response to Message 96637.  
Last modified: 20 May 2020, 9:14:24 UTC

The typical host averages of the 4-CPU computers is 3400 but the laptop is only 1800. With double the number of processors why is the 8 CPU host average only ½ of the four CPU computers?
What are your computing preferences set to? In particular-

Use at most xx % of the CPUs- should be 100%
Suspend when computer is on battery- if selected & the laptop isn't always on mains power, then no Rosetta work will done till mains power is back again.
Suspend when computer is in use- if selected, any activity on that system will stop Rosetta processing.
Suspend when non-BOINC CPU usage is above --- % If this is a low value (such as less than 85%), and other programmes are making use of the CPU (even just scripts on a web page) then Rosetta won't be able to do any work.

And as Tomcat suggested, it could be thermal throttling, and given that CPU's very low clock speed, any sort of throttling would significantly limit it's work output.
A programme such as CPUID HWMonitor will let you see if thermal throttling is occurring.
Grant
Darwin NT
ID: 96650 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Ed Machak

Send message
Joined: 10 Nov 16
Posts: 7
Credit: 17,339,411
RAC: 0
Message 96667 - Posted: 20 May 2020, 17:35:49 UTC - in response to Message 96650.  

Hello,

First I apologize for playing fast and loose carelessly interchanging the terms "cores" and "processors".

Thanks for the tips as they pointed me to thermal throttling. Unlike the four tower computers where the CPU operating temperatures run at about 75C in the laptop it reaches 100C. There's supposed to be a way to increase the laptop fan speed via a bios setting though I'm still looking for it :-)

Again, thank your for the prompt and detailed assistance.

Ed Machak
ID: 96667 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote

Message boards : Number crunching : Differences in host average between 4 and 8 processor computers



©2024 University of Washington
https://www.bakerlab.org