mersenneforum.org

mersenneforum.org (https://www.mersenneforum.org/index.php)
-   Information & Answers (https://www.mersenneforum.org/forumdisplay.php?f=38)
-   -   What are optimal settings for Prime95 Torture Test (https://www.mersenneforum.org/showthread.php?t=11674)

Discobadger 2009-04-01 11:40

What are optimal settings for Prime95 Torture Test
 
Hi,

I was hoping somebody can point me in the right direction.

I have a number of systems ranging from

System 1

* P43 Board
* E5200 2.5GHz 2mb L2 Cache @ 3.75Ghz
* 4GB 8500 Mhz Memory
* Vista 64Bit


System 2

* P45 Board
* Q9650 3.0Ghz 12mb L2 Cache @ 4.00Ghz
* 4GB 8500 Mhz Memory
* Vista 64Bit

System 3 (My latest beauty :) )

* X58 Board
* Core i7 920 2.66Ghz 8mb Cache @ 3.8Ghz
* 6Gb 1600 Mhz Memory
* Vista 64 Bit


I have had these system running for some time 90% stable and have benched them at what I presumed was the correct settings for Prime.

My routine is to normally build my system, flash to latest bios then Memtest using Memtest86+ for at least 10 full passes. This is carried out at standard settings to make sure I don't have any faulty hardware to start with.

After this is complete the Operating system, latest drivers and patches are applied.

I then set about overclocking :D

In the past I have just used Prime95 under torture test at standard settings using "In-place Large FFT's"
Normally the systems can run for 2 days quite easily without any errors.

However after speaking with a friend of mine he added that it will not be running a Complete system level test, it will only be torture testing the CPU.

Is this correct?
Should I be using small, blend or even a custom test?

Even though the system is running at 100% load the system is still pretty responsive, which is whats leading me to the assumption that I've not got prime setup & torture testing to the best of its abilities.

Regarding "options>cpu settings and information" Should I have a maximum memory amount set here?

If i was then to run a second set of tests with 3DMark 06 or 3DMark Vantage in the foreground to test VGA stability together with a CPU/Memory stress test I'd need to reserve approx 1GB of memory free. How would I set Prime to reserve at least 1024mb of memory.

What settings should I be using to give an overall complete system test?

Any help in clarifying these issues would be very helpful.

Many thanks.

cheesehead 2009-04-01 23:36

I'm not sure I'm the best qualified to make comprehensive specific recommendations for you, but here are a couple of responses about points I think I know: :smile:

(I presume others will correct me where I've grown rusty.)

[quote=Discobadger;167559]Hi,[/quote](Love your nym, BTW.)

[quote]In the past I have just used Prime95 under torture test at standard settings using "In-place Large FFT's"
Normally the systems can run for 2 days quite easily without any errors.

However after speaking with a friend of mine he added that it will not be running a Complete system level test, it will only be torture testing the CPU.

Is this correct?[/quote]Not really. "In-place Large FFTs" also well-exercises the memory controller and bus, though it doesn't exercise as large a range of RAM addresses as Memtest does.

[quote]Should I be using small, blend or even a custom test?[/quote]They each have their strengths/weaknesses for the purpose of torture testing (that's why there are choices, of course :-), and a comprehensive testing scheme would use each of them in turn.

The small-FFT test exercises the CPU (especially the FPU) and L1 cache hard, but puts only a modest load on memory buses and RAM other than L1.

The large-FFT test exercises the CPU (especially the FPU) and L1/L2 cache hard (but possibly not as hard as small-FFT, if memory fetch is slow) [I]and[/I] puts a heavy load on memory buses and non-cache RAM. However, the load it puts on main memory is not the same pattern as the load that Memtest puts on memory, so it is not a replacement for Memtest. Prime95 large-FFT is always fetching several megs of memory as fast as possible (thus stressing the memory controller and bus and cache), but is not systematically testing different parts of your RAM as Memtest does. (That's because the Prime95 torture test was not designed to be a general-purpose memory diagnostic test; it just turned out to be a good test in certain regards because the program optimizations are tuned to make maximum use of FPU and memory for purposes of fast FFT computation.)

The blend test steps through both small and large FFTs in a cycle, resulting in a combination of the small- and large-FFT tests.

The custom test is intended to allow specifying FFT size more particularly, in order to concentrate testing on a suspected critical size or range of sizes.

[quote]Even though the system is running at 100% load the system is still pretty responsive, which is whats leading me to the assumption that I've not got prime setup & torture testing to the best of its abilities.[/quote](1) Hey! Prime95's purpose is not to bring your system to its knees! Its purpose is to maximize throughput of certain mathematical computations (which might happen to inspire kneebends on some marginal systems, but that isn't the direct intent) while affecting responsiveness [I]as little as possible[/I].

Prime95 is not some random video game! It's serious (but not too-serious :smile:; maybe [I]sincere[/I] is more accurate) sincere software, carefully tuned over several years.

(2) Prime95 [U]is[/U] specifically designed to [I]interfere as little as possible with other tasks running on your system[/I]. It does this by running all of its computation-intensive code at the lowest system priority, so that any other task your system is doing can easily interrupt that part of the code when necessarily. So you can see 100% CPU, but you don't see that it's 99% at lowest priority and only 1% at higher priorities. That way, your system remains responsive to non-Prime95 tasks. [I]We're trying to be modest and unassuming while conducting blazing-fast calculations on enormously large numbers!!![/I]

Having said that ... when the large-FFTs are running, there may be a greater chance that other system tasks using large amounts of RAM will encounter higher swapping/paging rates than they would if prime95 were not running, because of somewhat increased competition for real memory space. Prime95's low priority will mean its pages will be the ones selected to page/swap out, but that does take a bit of time, plus the extra time to read in pages for the other tasks.

[quote]Regarding "options>cpu settings and information" Should I have a maximum memory amount set here?[/quote]Only if you'll do a lot of P-1 or ECM work. It's not documented well, but "available memory" matters _only_ during P-1 stage 2 and ECM stage 2. At all other times, such as during LL, DC or TF, it is ignored and doesn't matter.

cheesehead 2009-04-02 23:47

Oh, I [i]knew[/i] I was forgetting stuff ...

The large-FFT test is also the one that will make your system run hot if its cooling system isn't adequate, because it's the one that makes the highest percentage of your system's circuitry busy simultaneously.

I think you could use custom test and specify a very large FFT size and large memory to maximize the heat factor. (Someone correct me if I'm wrong here, please.)

So, since heat (i.e., inadequate cooling) is one of the most common reasons (perhaps [U]the[/U] most common reason) for a system to fail the torture test:

if you want the one single simplest possible torture test to maximally stress your system, do the large-FFT (or custom with specified large FFT/memory) test for 48-72 hours, so that your system goes through at least a couple of full daily temperature cycles in whatever room it's in.

Mr. P-1 2009-04-03 11:48

[QUOTE=cheesehead;167694]Only if you'll do a lot of P-1 or ECM work.[/QUOTE]

You won't get P-1 or ECM assignments even if you ask for them [i]unless[/i] you have more than 200MB available.

The only other assignment type which [i]may[/i] lead to stage 2 memory use are first time LL tests. This is because the separate P-1 effort has only just started, and most first time tests currently being assigned have not had this done.

Therefore, if you are going to do first time tests, it's worthwhile allocating at least 200MB, even if it's only for a few hours per day. It's not fatal if you don't, but you may end up expending a lot of CPU time on an LL test when a factor could otherwise easily have been found.

If you [i]are[/i] able to allow the client to use plenty of memory - at least 200MB, but 500MB is preferred and even more is better yet - then consider setting one core on P-1 assignments. The project needs more of this kind of work.


All times are UTC. The time now is 11:51.

Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2021, Jelsoft Enterprises Ltd.