Hyperthreading & Number of Cores. Parallel computing toolbox
Show older comments
I am using:
MATLAB Version: 8.3.0.532 (R2014a)
MATLAB License Number:
Operating System: Microsoft Windows 8.1 Version 6.3 (Build 9600)
Java Version: Java 1.7.0_11-b21 with Oracle Corporation Java HotSpot(TM) 64-Bit Server VM mixed mode
----------------------------------------------------------------------------------------------------
MATLAB Version 8.3 (R2014a)
Parallel Computing Toolbox Version 6.4 (R2014a)
I check my number of cores;
feature('numcores')
MATLAB detected: 4 physical cores.
MATLAB detected: 8 logical cores.
MATLAB was assigned: 8 logical cores by the OS.
MATLAB is using: 4 logical cores.
MATLAB is not using all logical cores because hyper-threading is enabled.
ans =
4
I also type;
>> p = parpool(8)
p.NumWorkers
Error using parpool (line 99)
You requested a minimum of 8 workers, but the cluster "local" has the NumWorkers property set to allow a maximum of 4 workers. To run a communicating job on
more workers than this (up to a maximum of 512 for the Local cluster), increase the value of the NumWorkers property for the cluster. The default value of
NumWorkers for a Local cluster is the number of cores on the local machine.
How can I allocate all 8 logical cores for matlab to use when running parfor.m for example? i.e. how can I make MATLAB use/ not use hyperthreading?
thanks
Accepted Answer
Jill Reese
on 9 Apr 2014
Edited: MathWorks Support Team
on 1 Sep 2022
10 votes
From the MATLAB desktop Parallel menu, select Create and Manage Clusters (or if you are using MATLAB version R2018a or earlier, select Manage Cluster Profiles). This opens the Cluster Profile Manager window. Select a local cluster profile, click the Edit button, and change the value of the NumWorkers property.
9 Comments
Matlab2010
on 10 Apr 2014
Edited: Matlab2010
on 10 Apr 2014
Jill Reese
on 10 Apr 2014
There is only a one-way relationship between
feature('numcores')
and the local cluster profile. The cluster profile picks up its default value from the number of physical cores in your system, as evidenced by the parpool message you quoted in your question: "The default value of NumWorkers for a Local cluster is the number of cores on the local machine."
The error message you quote also states that you can increase the value of NumWorkers for the local cluster profile up to a limit of 512. My answer tells you how to do this.
This change will allow you to create a local parallel pool that has more workers than physical cores in your system. This change only impacts parfor, spmd, and other explicit parallelism features that are found in the Parallel Computing Toolbox.
This change does not update the output of
feature('numcores')
which provides information about the physical hardware you are running on. Your physical hardware has not changed, but you have changed how you are instructing the Parallel Computing Toolbox to use that hardware when creating a parallel pool using a specific cluster profile.
I hope this helps to clarify the relationship between these values.
Jill Reese
on 10 Apr 2014
You only mention that the output of
feature('numcores')
is the same after you have increased the NumWorkers property to 8. Have you tried opening a parallel pool with 8 workers?
p = parpool(8)
This should work now, and I believe that this is what you were trying to accomplish, correct?
Jill Reese
on 10 Apr 2014
I went back and re-read your question, and I see now that the validation is getting caught up in validating required files. What are your settings for the three fields in the "Files and Folders" section of your local cluster profile?
Jeff L.
on 3 Jun 2014
I am having what I believe is a similar problem, but the above resolution hasn't helped me.
I have a 12 core Mac Pro with Hyperthreading, running R2014a. I see similar output from
feature('numcores')
but with 12 physical cores and 24 logical cores. The bottom line issue is that when I run a project with parfor (on the 'local' cluster pool), my CPU load is only 50%, and only 12 of the 24 logical cores are loaded. I have tried the above suggestion of
p=parpool(24)
with no success. Similarly, I have tried setting NumWorkers to 24 through a variety of approaches, and none work. How can I fully use my local machine's CPU (all logical cores)?
BTW, when I try to validate after setting 24 NumWorkers in Cluster Profile Manager, it fails to validate (12 works fine though).
Thanks for any help!
Ken Atwell
on 4 Jun 2014
Even if you could bump the core count from 12 to 24, I doubt MATLAB would see any benefit. Hyper-threading may help improve the efficiency of "light-weight" processes (email, ...), but MATLAB is typically compute-bound and really needs a full core. I too run a Mac with PCT routinely, and "50% utilization" is expected.
Here is a good discussion on why hyper-threading typically does not help MATLAB:
Thanks Ken, that link does discuss this quite thoroughly, and I found it helpful. Sounds like I should stop this quest for this "mythical" missing 50% cpu load. It seems that the way that the OSX Activity Monitor does its accounting is pretty misleading about the actual CPU load (misleading at least to those like me).
Seemingly corroborating this, I just tried turning off the hyperthreading on the Mac, and while the Activity Monitor showed 100% load on all the physical cores, my Matlab process run time was essentially identical to when hyperthreading was turned on (reporting a 50% CPU load).
Jeff L.
on 4 Jun 2014
To further follow up and put another nail in the coffin of my question, I did an additional test on a Mac laptop with 4 physical cores and 8 logical cores. When I ran my program (using parfor), it used 4 workers and took ~200 secs to run. The OSX Activity Monitor utility reported 50% CPU usage (essentially 4 of the 8 cores were used according to the monitor).
Then, I opened a second instance of Matlab on the same computer, and ran the program simultaneously in BOTH instances of Matlab. This time, the OSX Activity monitor reported 100% CPU activity and all 8 cores were fully used. HOWEVER , the run time for each instance was now 400 seconds. I.e., it took twice as long to run two instances of the program. This anecdotally shows that when running only one instance, it really is using close to 100% of the cpu resources, despite what OSX Activity Monitor says.
I hope this saves someone some time in the future. I wasted a full day trying to milk more performance out of my system.
Kay Gemba
on 19 May 2016
Jeff, the 'additional' 4 logical cores share their resources with the actual 4 cores. Hence using 8 will not help when maxing out the resources. Manufactures thus can then sell you '8' cores ;)
~kai
More Answers (1)
John Videtich
on 13 Apr 2018
3 votes
To resurrect an old thread, I ran a test on a machine with 24 cores / 48 logical (Xeon Gold 6146 x 2).
With 24 workers my calculation was taking 25.5 seconds. With 36 workers it was down to 21 seconds. With 40 workers it was down to 20.3. With 46 workers it was down to 19.8 seconds.
So it would seem hyperthreading does provide a real benefit - in this case, 22% reduction. I'm sure it's workload-dependent.
1 Comment
Shifan Gao
on 8 May 2018
Add on: my task takes 33 h with 4 pools and 19 h with 8 pools (i7 4 core). So new edition gives support to that kind of setting?
Categories
Find more on Parallel Computing Fundamentals in Help Center and File Exchange
Community Treasure Hunt
Find the treasures in MATLAB Central and discover how the community can help you!
Start Hunting!
