Hyper-Threading: Yes or No?

R

RFGuy_KCCO

DCOTM x4, [H]ard|DCer of the Year 2019
Joined
Sep 23, 2006
Messages
923
What's the consensus for DC'ing? I have kept it active on all my crunchers, but is it better to turn it off? What are the cons to leaving it on, if any?

It appears WU's for most projects take ~75% longer with HT on, than they do with it off. Therefore, I believe I would get more WU throughput over time with it left on. However, it's possible I am missing something.

Would like to hear the opinions of my team mates, since many of you have been doing this much longer than I have. Thanks!
 
Thanks! Looks like most projects do better with it on, except for Primegrid. Interesting, for sure. I wonder if it because all the latest LLR apps for all their sub-projects use AVX?

Oddly, I did a search in this forum and came up with nothing. I knew there had to be something here about this.

On WCG Mapping Cancer Markers WU's, I get the following results:

HT on: ~5:45:00 per WU
HT off: ~3:30:00 per WU
 
Last edited:
With a perfectly threaded program ( no locks in the code ), you will gain a max of about 20% -25%speed increase with HT. This is taking into account that the program is also designed to not cause cache thrashing.

You also have to take into account the L1 and L2 cache that will be shared between 2 threads per core with HT on, and that the L3 cache is also shared between all the threads.
 
RFGuy_KCCO said:
Interesting, for sure. I wonder if it because all the latest LLR apps for all their sub-projects use AVX?
Click to expand...

Speaking of AVX, how many projects out there, besides some of the Primegrid stuff, benefit from CPUs that have AVX or AVX2? Crunching one of the recent challenges was quite frustratingly slow for me as none of my CPUs I have support AVX.
 
I know that Einstein does, though they have made the workunits larger to accommodate the increased processing ability. Same thing for their GPU apps, once they were able to optimize properly, they became longer and larger. So speed may not be an indicator sometimes.
 
Asteroids has an AVX version of their CPU app. If you use the Lunatics Optimized Apps for SETI and SETI Beta, they also have an AVX enabled version for both MB and AP WU's.
 
cyclone3d said:
With a perfectly threaded program ( no locks in the code ), you will gain a max of about 20% -25%speed increase with HT. This is taking into account that the program is also designed to not cause cache thrashing.
Click to expand...
Locks in code has less to do with performance increase than filling available spots in the execution pipeline for HT performance.

The 20-25% increase often seen in speedups of general code is generally due to holes generated by cache misses and not having enough work waiting (due to type of code or dependencies) at port reservation stations. The upper limit of a speed up for pathological code can be much more than 25%, or a slightly negative in other cases. It is totally dependent on the code being used, but later iterations of HT (versions released after Core) are much better at preventing performance hits, at least if the OS is aware that HT != 2 full cores (i.e. high utilization threads should be distributed to different cores when possible).

An easier way to look at it is that typical throughput on modern processors on general code is 1.x instructions per core per clock, when the theoretical maximum is often 2-3x higher. It's all about filling underutilized resources in the execution pipeline, which is also why some code sees little to no improvement.
 
tjmagneto said:
Speaking of AVX, how many projects out there, besides some of the Primegrid stuff, benefit from CPUs that have AVX or AVX2? Crunching one of the recent challenges was quite frustratingly slow for me as none of my CPUs I have support AVX.
Click to expand...

It appears that it may not really matter anyway tj it appears that Boinc point system may be predigest towards AMD anyway. My AMD rigs are faster at Numberfields than my Intels but the Intels get 30% more PPD than the AMD's :( The project scientist is looking into it now so we shall see.

http://numberfields.asu.edu/NumberFields/forum_thread.php?id=209
 
I know the issue between AMD and Intel has been brought up at WCG before, but I don't think there was any official response or solution on the matter.
 
You must log in or register to reply here.
Back
Top