LLR beta Version 3.8.13 (deprecated)

[KEP]

Originally Posted by AG5BPilot

Yes, you're confused about the debate at PrimeGrid. :)

We're not concerned about the validity of any of the BOINC jobs run with 3.8.12. The discussion I think you're referring to was about the qualification testing we're doing for 3.8.13, which must be done with 3.8.13. Someone essentially asked if he could use a test performed with 3.8.12 for the 3.8.13 qualification, and my answer was no. But BOINC tests performed with 3.8.12 are ok if they pass validation.

Well I was confused regarding the discussion in this thread. But I'm glad that more and more, besides Jean has confirmed that I shouldn't suspect any problems with my currently ~40000 tests, wich I also believe is the equavilent of not having missed any primes

Thanks for all your reassurances

Take care

Kenneth

[pepi37]

When I do single test llr64 -d -q"candidate" and that candidate is prime, with BeepOnSuccess=1 there is no sound.

But if I make test with llr64 -d -a input file and there is prime in that sequence then I in my speakers hear beep beep beep :)

So is this feature or bug? :)

Win7 x64
llr64 bit 3.8.13 cli

[Citrix]

On a Haswell i7 FMA3 AVX2, I am getting 100% benefit (time has halved) with 2 cores running and 25% benefit with 4 cores. Is the benefit this significant only for my computer?

[kracker]

Originally Posted by Citrix

On a Haswell i7 FMA3 AVX2, I am getting 100% benefit (time has halved) with 2 cores running and 25% benefit with 4 cores. Is the benefit this significant only for my computer?

What memory do you have? Depending on the workload, memory makes all the difference.

[Citrix]

8Gb RAM
L1 32Kb
L2 256Kb
L3 6MB

[kracker]

Originally Posted by Citrix

8Gb RAM
L1 32Kb
L2 256Kb
L3 6MB

What I need to know is the memory speed/channel.

[Cruelty]

Originally Posted by kracker

What I need to know is the memory speed/channel.

+ timings (best if you post CPU-Z screen)
BTW: some nice articles on memory scaling Ivy Bridge and Haswell.

[kracker]

Ah good idea.

Is LLR affected by latency? I know Prime95 is not/minor.

[KEP]

Originally Posted by Citrix

On a Haswell i7 FMA3 AVX2, I am getting 100% benefit (time has halved) with 2 cores running and 25% benefit with 4 cores. Is the benefit this significant only for my computer?

On my Haswell i5 4670 / 3.4 GHz Processor running with DDR3 1600 MHz PC3-12800 RAM, I'm seeing (running all 4 cores) a 16% speed increase on base=383 tests, compared to LLR version 3.8.11. All test is run using the 64-bit version of Cllr64. So a pretty need increase in speed. I'm not sure what the impact will be if I run on only 3, 2 or 1 core. My previous experience from the time when George first upgraded the GWNum library to support faster FFTs for k*b^n+/-1 tests, is that it is generally a little faster to run tests on all 4 cores since my productivity will be a few percent higher. Weather it be true with Haswell is really not something I've checked in to.

[kracker]

Originally Posted by KEP

On my Haswell i5 4670 / 3.4 GHz Processor running with DDR3 1600 MHz PC3-12800 RAM, I'm seeing (running all 4 cores) a 16% speed increase on base=383 tests, compared to LLR version 3.8.11. All test is run using the 64-bit version of Cllr64. So a pretty need increase in speed. I'm not sure what the impact will be if I run on only 3, 2 or 1 core. My previous experience from the time when George first upgraded the GWNum library to support faster FFTs for k*b^n+/-1 tests, is that it is generally a little faster to run tests on all 4 cores since my productivity will be a few percent higher. Weather it be true with Haswell is really not something I've checked in to.

1600 will heavily handicap performance if it is similar to P95. Hey, do us a favor and compare 1 core/4 cores
If it is not bottlenecked, one core should be similar to running four.

[AG5BPilot]

It just so happens that I have an i5 4670K with 2400 MHz memory. A week or so ago I ran tests to see how much of an impact (loss of efficiency) I experienced going from 1 to 4 cores, first with the memory at 1600 MHz and then with it at 2400 MHz. Here's a copy of my original post at PrimeGrid:

Cores	core 1	core 2	core 3	core 4	Iter/sec	I/S/core	Incremental	Loss
1	8.515				117.4398121	117.4398121	n/a        	0.00%
2	8.761	8.763			228.2583915	114.1291957	110.8185794	2.82%
3	9.495	9.488	9.509		315.8784081	105.2928027	87.6200166	10.34%
4	11.073	11.896	12.001	11.056	348.1466467	87.03666167	32.2682386	25.89%

I increased the memory speed to 2400 MHz. I *think* the BIOS, on its own, increased the CPU base speed from 3.4 GHz to 3.6 GHz. I'm not totally sure why, or if I'm understanding everything correctly. Therefore, some of the raw speed increase might be due to the CPU simply running faster. However, it's clearly visible in the results how much better the memory does at 2400 MHz because the difference between 1 core running and 4 cores running is substantially smaller.

Cores	core 1	core 2	core 3	core 4	Iter/sec	I/S/core	Incremental	Loss
1	7.893				126.6945395	126.6945395	n/a        	0.00%
2	8.065	8.066			247.9697486	123.9848743	121.2752092	2.14%
3	8.336	8.350	8.334		359.7124921	119.904164	111.7427435	5.36%
4	9.248	8.882	9.244	9.048	439.4186831	109.8546708	79.70619097	13.29%

Getting faster memory (assuming your MB and CPU support it) is clearly a very cost effective means of getting more LLR performance. (In my particular purchases, with Internet price fluctuations being what they are, the 2400 MHz memory was actually $1 cheaper than the otherwise identical 1600 MHz part.)

Temps are running about 5 degrees higher, but the fans are only running at about 60% speed. We'll see if I get any LLR tasks fail validation.

As you can see by the "loss" column all the way on the right, the loss due to memory contention with 4 cores running drops from 26% at 1600 MHz ro 13% at 2400 MHZ. (Due to the BIOS overclocking the CPU to 3.6 GHz when using the XMP profile, which would cause even more memory contention, the benefit of the faster memory is actually slightly better than the numbers indicate.)