Cores or Speed

Lester Lane
Lester Lane
Joined: 30 Aug 05
Posts: 10
Credit: 211011021
RAC: 0
Topic 207442

Hi all,

If I'm not mistaken I think Einstein works better/faster if the CPU has more cores, rather than simply a fast chip.  Am I right?  Is the app heavily multi-threaded and needs a high number of cores, or is it not so dependant of cores and I should buy a fast chip?  Are most of the other BOINC apps written this way or is Einstein the exception?

Thanks in advance.

LL

archae86
archae86
Joined: 6 Dec 05
Posts: 3157
Credit: 7231874682
RAC: 1162902

I don't believe any of the

I don't believe any of the current Einstein applications are meaningfully multithreaded.  The CPU-only tasks will happily use very nearly the full capability of a single CPU (whether a hyperthreaded representation or not) if system scheduling permits.  The GPU support tasks are less greedy in the sense that most of them won't keep the CPU they are running on fully busy under most circumstances.  Depending on application, GPU, and CPU, some have kept a host CPU occupied at quite a low rate.  Nevertheless, the productivity of the GPU can be heavily influenced by latency (how long after the GPU requests CPU service does a CPU actually get going on the task?).  Hence many of us take measures for the sake of improved GPU productivity for which the means to practical benefit is a reduction in CPU latency on these requests.  The most common measure is simply reducing the system number of BOINC CPU tasks allowed to run, though other measures include not running BOINC CPU tasks at all, or altering the scheduler behavior by altering default affinities and priorities--for example by using Process Lasso).

You do get benefit from lots of cores in running Einstein, but the mechanism is not multi-threading on a single copy of the application, but rather just by running more simultaneous copies.

As a first order rule your Einstein productivity will be driven by the product of hour number of cores times the capability of each core.  Obviously there are further complications such as competition for shared resources in particular memory access.

AgentB
AgentB
Joined: 17 Mar 12
Posts: 915
Credit: 513211304
RAC: 0

Just a small addition, use of

Just a small addition, use of quad channel memory (increases memory bandwidth) does produce a benefit especially when the number of cores / threads increase.   This typically shows up in CPU tasks with large memory requirements.  If you see 1 task taking N seconds per task, then a 4 core should run 4 tasks in each close to N seconds.    If you see performance drop after running 2 tasks then you might want to look at this. 

Some very good material here https://software.intel.com/en-us/articles/intel-performance-counter-monitor/

Lester Lane
Lester Lane
Joined: 30 Aug 05
Posts: 10
Credit: 211011021
RAC: 0

Many thanks, I'll increase

Many thanks, I'll increase the number of cores and see if I can include this quad channel memory too.  Within budget of course!!

Lester Lane
Lester Lane
Joined: 30 Aug 05
Posts: 10
Credit: 211011021
RAC: 0

Oh, by the way...Which will

Oh, by the way...

Which will perform better: Intel® Xeon® Processor E5-1680 v4 or Intel® Xeon® Processor E5-2699A v4, and why please?  And I plan on using the NVIDIA GTX 1080 GPU.

Thanks again,

James

MarkJ
MarkJ
Joined: 28 Feb 08
Posts: 437
Credit: 139002861
RAC: 0

Lester Lane wrote:Oh, by the

Lester Lane wrote:

Oh, by the way...

Which will perform better: Intel® Xeon® Processor E5-1680 v4 or Intel® Xeon® Processor E5-2699A v4, and why please?  And I plan on using the NVIDIA GTX 1080 GPU.

Thanks again,

James

The E5-2699A would be faster (22 cores/44 threads) at 3.4Ghz. Its got a larger cache. The list price is almost $5000 USD. You might be better off getting multiple machines with a lower core count. if you can wait for the AMD Naples its got 32 cores/64 threads. It also supports multiple CPU's per motherboard like the Xeon.

I have a couple of AMD Ryzen's being built (8 cores/12 threads) about a tenth of the price of the Xeon. I also have 8 i7-6700 machines (4 cores/8 threads).

Gary Roberts
Gary Roberts
Moderator
Joined: 9 Feb 05
Posts: 5872
Credit: 117838341690
RAC: 34783086

Lester Lane wrote:Oh, by the

Lester Lane wrote:

Oh, by the way...

Which will perform better: ...

You need to tell us more about your objectives.
  - do you have a particular search you're interested in, or do you want to support all searches?
  - are you considering 'right now' or are you more interested in performance in the longer term?
  - is the machine for 'crunching only' or does it have s different primary purpose?
  - do you see yourself still crunching in the longer term?
  - is energy efficiency important to you?

Depending on your answers to these and other considerations, what people might recommend could easily change.

If you were most interested in the continuous GW search and your interest was 'right now' then raw CPU power and lots of cores would maximise your output, but at considerable cost.  If you want to take a longer term view, CPU power will lessen in importance and energy efficient GPUs will become increasingly important.

So, spell out in detail what your objectives are.  Think it through and don't be in too much of a rush :-).

 

Cheers,
Gary.

Lester Lane
Lester Lane
Joined: 30 Aug 05
Posts: 10
Credit: 211011021
RAC: 0

Thanks for the feedback. I

Thanks for the feedback. I used the £5k chip as an extreme example to help me understand what helps. I like the AMD route but there is no workstation board yet, I believe, indeed there may never be. I've blown my current PC burning BOINC apps with Gigabyte board so now I'm going serious! What board are you using for your AMD chip please and why?

i am currently banging away at GPUGRID, Asteroid, SETI, Milkyway, Enigma, and opted for all feeds from Einstein. I get some for GPU (GPB) and some for CPU (CGW) 

I am looking to build now and only a little green

mikey
mikey
Joined: 22 Jan 05
Posts: 12709
Credit: 1839113724
RAC: 3643

Lester Lane wrote:Oh, by the

Lester Lane wrote:

Oh, by the way...

Which will perform better: Intel® Xeon® Processor E5-1680 v4 or Intel® Xeon® Processor E5-2699A v4, and why please?  And I plan on using the NVIDIA GTX 1080 GPU.

Thanks again,

James

I believe the 1080 will require a new "uefi" bios, as my 1060 does, and won't work in older machines so be careful if you are buying used. I had planned to put my 1060 into an older pc and after a few days of it 'just not working' I found a forum where they explained that. It was a Dell forum as the pc was a used Dell, the best gpu I could put in the pc was a 750Ti so I did and it works just fine. The 1060 works just fine in an AMD 6 core pc that does have the newer "uefi" bios in it.

mmonnin
mmonnin
Joined: 29 May 16
Posts: 291
Credit: 3435756540
RAC: 4128005

I just put a RX 580 into an

I just put a RX 580 into an EVGA 680 SLI motherboard. Its old and still runs a new generation of GPU just fine.

 

I think that problem is due to it being a Dell and probably wasn't able to power a 1080 irrelevant of the BIOS.

Robert Klein
Robert Klein
Joined: 2 Dec 16
Posts: 11
Credit: 1886554656
RAC: 8347

For E@H a

For E@H a single-precision-monster is needed.

Maybe the following config gives a strong basis for relatively low price and energy consumption and gives much space for upgrades in the future:

 

AMD Ryzen 5 1600X (270 Euro)
2x AMD RX 580 (2x 270 Euro)
Gigabyte Aorus GA-AX370-Gaming K5 (170 Euro)
2x 8GB DDR4-3200 RAM (150 Euro)
Intel SSD 600p 128GB, M.2 (60 Euro)
Corsair RMx Series RM850x 850W ATX 2.4 (125 Euro)
Corsair Carbide Series Quiet 600Q (140 Euro)
Thermalright True Spirit 140 Power (50 Euro)

 

Total cost: 1.500 Euro; 13.000 GFLOPs; 600W power-consumption.

Could be upgraded to 2x 1080Ti: 24.000 GFLOPs

Comment viewing options

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