27.7 vs 27.9 for Stress Testing?

[Belial88]

i7-3770K (with hyperthreading, so i run 8 threads naturally)
Z77X-UD5H
2x2GB XDZ PSCs 2400 CL8
gtx 460, cx500, source 210, nh-d14

so is the new version better for stress testing? I am a religious believer in 24+ hours of prime95 for stability testing everything, and that is the easiest and quickest way to find stability (gave up ibt, occt, hyperpi, memtest, etc, when faulty hardware and way off overclocks could pass 12 hours on all of those programs, even 12 hours of p95).

Thanks. It'd also be cool if like 12 hours of the new version = 24 of old or something? Man i wish you could just start blend tests where they left off or something ;/

[kracker]

I believe both are same for stress testing, 27.9 fixes a bug on running more than one core(thread) on a worker, I believe it does not affect stress testing.
And I don't think there is a point on resuming or anything like that, I think the tests just loops.

[Belial88]

Blend loops over ~23 hours though... that's why I wish you could start it from a certain fft length, pause it, etc.

, 27.9 fixes a bug on running more than one core(thread) on a worker, I believe it does not affect stress testing.

I don't understand what you mean. Like at all. I have an i7, the default is 8 worker threads, for stress testing.

On the main page it says something about improving the AVX for sb/ib?

This latest version of prime95 has been optimized for Intel's new AVX instruction set. Owners of Sandy Bridge and Ivy Bridge CPUs (Core i3/i5/i7 2xxx and 3xxx) will see a substantial performance boost.

[kevindd992002]

Belial, I take it that you are the same Belial at OCN?

I suppose you are using P95 27.7 build 2 right now, yes?

[Belial88]

yesyes

[LaurV]

You can use your cores in more (different, alternating) ways:
(a) with HT disabled: you use 4 physical cores, you can have either
(a1) 4 independent workers, each one single-threaded,
(a2) 2 independent workers, each one with 2 threads
(a3) 1 worker, taking all 4 threads
(b) with HT enabled:
(b1) 8 independent workers, each one single-threaded,
(b2) 4 independent workers, each one with 2 threads
(b3) 2 independent workers, each one with 4 threads
(b4) 1 worker, taking all 8 threads

This you can set from the "Test/Worker Windows" menu, and is taken into account during the stress test. I suggest you do "useful stress testing", i.e. take some exponents from PrimeNet and do some DC, to which you can compare the final result with the known residues, you satisfy your need of stress testing, you help the project, you get credit too, and if you are lucky and find a missed prime, you may get some money too, like 3000 bucks or more.

For your CPU/system, the best "production" result you will get with case (a1).

Case (b2) will be more "stressful" for the system, but for production, the output will depend on many factors, like your memory speed, exponent tested, other running applications. What I want to say is that in (b2) case you may witness increasing or decreasing of the total output, comparing with (a1) depending of how fast you CPU can swap the caches (between P95 worker instances and other running applications). In all cases, you will witness a BIG increase in the consumption and produced heat with (b2) compared with (a1). Therefore, most people here agree that HT is not useful for LL/DC testing with P95. It brings too less benefit for the additional expenses it needs.

If you have a (really-)fast memory interface, and the computer is mostly idle (no big applications running to swap contexts most of the time) then (b1) might be better for "production" (about 5% more output), otherwise (b1) is limited by the memory access, and generally the "stress" is lower then (b2). No idea how the stress compares with (a1), it depends on your system.

You can try the "Options/Benchmark" and compare the numbers.

Toward the (a3) or (b4) case, they are good when you need to test one exponent fast, dividing the time by 4, respective 4.xx (smal xx depending on your system), but for long term running, (a1) is still the best (output, expenses, wear of the system in time, etc).

(hell, I said this story many times, can't we make a tutorial somewhere and just refer the links? )

Originally Posted by Belial88

I don't understand what you mean.

He says that the bug in v27.7 affected cases (b2), (b3), [and not sure if (b4)] which could occasionally produce wrong results. (the bug appeared for me when both "multi-threaded" "multi-worker" was on effect, and HT enabled, i.e. helper threads on logical cores). The stress testing was not affected.

[LaurV]

Just to be clear, "HT enabled" or "disabled" effectively refers to the number of workers, not to your bios/OS settings. P95 is enough clever to allocate the physical cores in an intelligent way, so if you set a total of less then 5 threads for it (like cases a1, a2, a3), he will always use physical cores. If you set 5 or more, he will use logical cores, IF they are enabled in your system.

[Belial88]

For your CPU/system, the best "production" result you will get with case (a1).

But worker windows already says 4 right now, which is the default. And then at default in torture test it says 8 threads. i get 8 windows during testing though.

take some exponents from PrimeNet and do some DC, to which you can compare the final result with the known residues, you satisfy your need of stress testing, you help the project, you get credit too, and if you are lucky and find a missed prime, you may get some money too, like 3000 bucks or more.

I just have the faintest clue that prime95 was originally designed a long time ago for finding prime numbers. I really havent an idea what you are talking about.

Your whole post pretty much lost me. I'm just trying to stress test, I don't do anything with finding prime numbers. I would like the most intensive stress program, if there's another way to test that is more likely to reveal instability than standard custom blend test (set 80%+ ram, set worker priority to 10/above normal in task manager), I'd love to hear it.

do some DC

Google tells me distributed computing is what you mean, but i have no idea what the context is or what you mean by this.

[LaurV]

Originally Posted by Belial88

And then at default in torture test it says 8 threads.

Whoops! My bad! I forgot you can select the number of threads there, but indeed you are right. To answer to your question, the one which started all this, there is no difference between the two versions of P95 program, related to stress testing.

To answer to your newest question, you may wonder to find out that 7 workers (or less) can stress your computer more then 8 workers can do, assuming you have some bottlenecks in your system where more workers will wait for each-other, but less workers will not. Like a slow memory access is the best example.

The best way to find out (which does not involve advanced knowledge) is to connect a power-meter to the electrical plug of your system (like a kill-a-watt or so) and see which testing mode takes more power. THAT will be the one which is stressing your system more. Well, this makes me think that a long discussion can be started from here too, regarding of what parts of the system you want tested, because this will arguably stress your power supply too, and if that is what you want tested also, then turning on/off different peripherals can do a better job, or not.

On the other side, P95 was never a good toy to test the memory or the memory interface, the VGA cards, the PCI-e slots, etc. What P95 is "genial to", is testing your CPU, and CPU-cooling assembly, under really hard conditions. It is the simplest and the best program if you want to find issues with your CPU, with your thermal dissipation system, or if you want to identify "on the edge" cooling or power supply problems. No other software beats it at this.

P.S.: by "DC" I was referring to "Double Checks", a kind of "test" we do here around, it is enough stressful for your hardware (beaten only by the LL test, but for LL test you don't know if your hardware was good or bad, unless someone else "Double Checks" your result, which can take months or years). One LL test can take around 7-20 days on your system (depends on exponent, number of threads, etc, even longer for exponents outside of our current testing range). One DC test can take about 2-7 days for your system.

[prgamma10]

Originally Posted by Belial88

Your whole post pretty much lost me. I'm just trying to stress test, I don't do anything with finding prime numbers. I would like the most intensive stress program, if there's another way to test that is more likely to reveal instability than standard custom blend test (set 80%+ ram, set worker priority to 10/above normal in task manager), I'd love to hear it.

I suggest GNFS (factoring a 130 digits number with all 4 threads is a good workout for your PC, especially the Lanczos phase).

[Dubslow]

GNFS stress testing takes much longer than a Prime95 stress test, and is much harder to do.

Version 27 uses AVX, which will increase the speed of production work done for GIMPS, and does not have any relation to how long a proper stress test should go.

27.9 fixes a bug in 27.7 where multithreaded workers (cases a2, a3, b2, b2, b4) would spuriously error out. Definitely use 27.9. (As with all programs, the latest is generally best.) Anyone who is using 27.7 (or any other version 27) should definitely switch to 27.9, especially if they are running more than one thread per worker.

You describe that the torture test defaults to the b1 case (8 separate windows), which I think is the most stressful: taxing the hyperthread context switching, and maximal memory bandwidth (for sufficiently large FFTs). Just leave it as it is and run it for the 24 hours that is agreed.

You can manually set what FFT lengths a torture test runs via one of the drop down menus; fiddle with it a bit.