3770K versus other cpus

Atomant
Atomant
Joined: 16 Mar 13
Posts: 3
Credit: 1129915
RAC: 0
Topic 196918

Hi All,

I am wondering why it takes longer for a I7-3770k cpu
than an i7-3610QM to complete Gravitational Wave S6 WUs.
I have seen other WUs like the following including my own (I7-3770k).

examples:
http://einsteinathome.org/workunit/158854760

http://einsteinathome.org/workunit/158934064

My I7-3770k is running without interruption and when compared to another 3770k the run times are about the same.

Found a problem with my system running the GPU BRP units, screen blanker, and the video card dropping the core clock for whatever reason. Fixed that issue. The BRPs seem to run fine between 880 and 890 seconds per unit now.

Any ideas why my run times are slower on the cpu tasks ?

Thanks Atom

archae86
archae86
Joined: 6 Dec 05
Posts: 3145
Credit: 7023204931
RAC: 1829022

3770K versus other cpus

It is a bit awkward to compare times among systems.

Some factors which can affect comparisons which are not captured by the CPU part number:

1. What clock rate is the part actually running at? (this is not captured on the BOINC system description information--which just accepts what is self-reported, so people who overclock but don't tamper with the self-reporting generally show a lower than actual clock rate).

2. If the CPU is HT-capable (both of the ones in your comparison are), is hyperthreading actually active? (generally this is a BIOS option--I think the default setup on most modern boards it to enable it). Some users prefer to turn HT off. While it generally provides higher total CPU throughput, the power efficiency of the increment over non-HT operation is poor, and extreme overclockers have reported that on some CPU/application combinations the maximum achievable overclock is enough lower to completely negate the HT throughput gains.

3. How many actual CPU tasks are active? This has special relevance for HT-capable CPUs, as people allowing only as many CPU tasks as physical cores (or even fewer) will get substantially shorter elapsed time, to a much greater degree than is true on non-HT hosts. Many people limit CPU tasks hoping to get greater total system throughput from higher GPU throughput, and some people limit CPU tasks for better power efficiency, or in hope of better system responsiveness.

4. Is any other form of Throttling active? Examples include the power-saving option provided by BOINC itelf, or Fred's TThrottle, and perhaps others. These generally lower throughput and extend elapsed times. People use them to avoid high temperature, lower power consumption, and perhaps other reasons.

5. Is there a GPU on the host running Einstein? This imposes support requirements which will lower the system throughput for CPU-only jobs.

6. Is a priority raising scheme in use to improve GPU job throughput by favoring the CPU support task for the GPU job more than the default? If so, this will, naturally, have the effect of extending the elapsed time reported on CPU tasks.

7. If a GPU running Einstein is present, how many simultaneous jobs are running on it? More simultaneous GPU jobs generally raises throughput at least up to two, and often up to three jobs. But the CPU and bus traffic support required rises much faster than the GPU throughput, so these higher simultaneous GPU job hosts will show lower CPU throughput.

If I looked up your part number selections on ark.intel.com correctly both are Ivy Bridge-based 4-core HT CPUs, so to first order, were all other things equal, one might expect similar clock counts for GW jobs execution. I've tried to suggest a few of the items which might alter things. I have no opinion as to whether most of them apply in your specific comparison.

Atomant
Atomant
Joined: 16 Mar 13
Posts: 3
Credit: 1129915
RAC: 0

RE: 2. If the CPU is

Quote:

2. If the CPU is HT-capable (both of the ones in your comparison are), is hyperthreading actually active? (generally this is a BIOS option--I think the default setup on most modern boards it to enable it). Some users prefer to turn HT off. While it generally provides higher total CPU throughput, the power efficiency of the increment over non-HT operation is poor, and extreme overclockers have reported that on some CPU/application combinations the maximum achievable overclock is enough lower to completely negate the HT throughput gains.

3. How many actual CPU tasks are active? This has special relevance for HT-capable CPUs, as people allowing only as many CPU tasks as physical cores (or even fewer) will get substantially shorter elapsed time, to a much greater degree than is true on non-HT hosts. Many people limit CPU tasks hoping to get greater total system throughput from higher GPU throughput, and some people limit CPU tasks for better power efficiency, or in hope of better system responsiveness.

Thanks for the great reply archae86. Looks like the answer is the HT.
Ran just one cpu task and nothing else; run time was about 3 hours instead of 4 1/2. At least I know now.

I've overclocked this system, but somewhat disappointed with the results.
CPU wise, I guess I'll keep crunching 8 every 4 1/2 hours rather than the 4 every 3 hours.

Thanks again.

Filipe
Filipe
Joined: 10 Mar 05
Posts: 175
Credit: 365520506
RAC: 45243

RE: I've overclocked this

Quote:
I've overclocked this system, but somewhat disappointed with the results.
CPU wise, I guess I'll keep crunching 8 every 4 1/2 hours rather than the 4 every 3 hours

Are you running more than 1 GPU task simultaneously? Whith a 670GTX, you can run 2 or 3 tasks simultaneously and have a greater RAC.

Also, on a HT capable cpu, you don't have to choose beetwen 4 or 8 cores running.
I think you will have faster results using 6 cores, and leaving the others 2 for feeding the GPU.

Filipe

Atomant
Atomant
Joined: 16 Mar 13
Posts: 3
Credit: 1129915
RAC: 0

I tried running more than one

I tried running more than one GPU task without any cpu tasks, but if I recall correctly, the GPU tasks took longer overall. I will attempt it again now that I fixed the GPU screen blanking issue and have a better understanding of the CPU/core/HT tasks.

Thanks.

Comment viewing options

Select your preferred way to display the comments and click "Save settings" to activate your changes.