About Collatz Conjecture

[TennesseeTony]

Main Page: http://www.thesonntags.com/collatz/index.php

Collazt is one of the possible projects for a future Sprint race in the Formula BOINC race, so I thought I might want to get familiar with it.

This private project has a real axe to grind with spammers, and doesn't allow you to simply join from the BOINC client, you have to join from the home page, and you have to have an invitation code. Additionally, users without credit will be deleted, and so will clients/computers without credit. After you join, you might as well go ahead and set your preferences (GPU only for example) before you add the project to your BOINC manager.

I have tried joining the project before, but never received my invitation code. So I tried again and now have it. It is spammerssuck and is subject to be changed without notice.

On the main page, (linked above at the start of this thread) you will need to carefully read the system requirements. What I missed was that you need to download and install both x86 AND 64-bit versions of C++, found here.

The instructions say it's needed for CPU crunching, but it is needed for GPU crunching as well, as I found out when all my tasks immediately showed computation errors on a GTX1080. (Then I got locked out of requesting more tasks for 24 hours, due to the errors. Just close and restart BOINC if that happens)

In all, a real pain compared to most projects, and it feels a bit unfriendly in the forums when other's also missed the info on the main page. I have no idea why anyone would want to run this project.....oh, I see, uh, got a couple tasks in already, now I know why they run it... 36,000 points every 13 minutes is 4,000,000 ppd per 1080!

More importantly, practically no CPU usage (0.7%).

So now that the introduction for the new guys (myself included) is out of the way, how about an experienced Collazt user tell us what app_config to use for certain cards? My 1080 was at 90% usage already, but the more results the better!

 

[crashtech]

If they make it too hard, they are going to miss out on well-meaning people that just don't have a lot of time to jump through hoops.

 

[crashtech]

Is there any point to using AMD cards, or would I simply switch my 1060s over from folding duty, IF, it's a part of Formula BOINC?

 

[TennesseeTony]

A set just finished up on the R9-280X's. Not a very efficient use of the cards, but not too bad either.

 

[crashtech]

A set just finished up on the R9-280X's. Not a very efficient use of the cards, but not too bad either.

Are those 1080 scores, there? Half the points, but quite a bit more CPU heavy, is that right?

Oops, I meant twice the time, anyway half as good lol

 

[TennesseeTony]

Yep, that's a 1080 versus R9-280X, same model CPU/MB, same speed RAM.

 

[crashtech]

That's not horrible, are you advising that we jump through the hoops just in case?
Edit: I've gone ahead and asked for an invite code.

 

[crashtech]

Well, got my invite code and I am lending a few cycles just to see how it goes. Probably need to ferret out an app_config example that can be edited to fit my needs.

 

[crashtech]

So I searched their forum for "app_config," no joy..

 

[TennesseeTony]

OCN has an extensive 51 page thread on Collazt. Some pages are dated, but from what I gather, you only need to run one task at a time, but you need to 'tune' your video card. The code below is inserted into your GPU.config file, same way you modify the app_config. Collazt uses both 32 bit and 64 bit, so there will be TWO files to modify (once your computer has run both 32 and 64 bit tasks).

verbose=1
kernels_per_reduction=48
threads=8
lut_size=17
sieve_size=30
cache_sieve=1
sleep=1
reduce_cpu=0

On a standard windows 10 installation these files can be found here:

Past my bedtime, so I have no idea what those do, except the kernels_per_reduction=48 affects how large a chunk of the work unit the GPU 'bites off' and tries to chew at once. It seems most folks were using this particular .config on 980Ti's, but also 7970/280X's, so I have no idea how it will work on other cards.

 

[TennesseeTony]

Apparently it works pretty good for a 1080, check the improvement in times (but a reduction in points?)

The 280X hasn't finished yet, but it also looks to have reduced it's time by 45% as well, it's past 50% in 7 minutes.

 

[crashtech]

My collazt project directory only has one .config file for AMD GPU, "collatz_sieve_1.21_windows_x86_64__opencl_amd_gpu.config" so I modified just the one.

 

[Orange Kid]

Just to let everyone know. The optimized .config files will reduce the amount of time to process the WU's and you will notice less credit per WU, but overall it will increase the credit per day.
There is also a thread on the Collatz forum on the subject

 

[iwajabitw]

Requested a code, will see about testing tonight when I get home.

 

[crashtech]

There is no actual app_config as I understand it? But you can edit files as Tony shows above.

 

[Orange Kid]

There is no actual app_config as I understand it? But you can edit files as Tony shows above.

Yes, you are correct.
Running more than one WU at a time is not recommended for this project.

 

[TennesseeTony]

Is the code in the first post not working? (tried to remember it after I activated my account and trashed the email)

Current Invite code: spammerssuck

 

[iwajabitw]

I got the same invite code as you Tony so it must be a generic for the time being.

 

[iwajabitw]

Account created and joined the team, after the formula race starts ill see about getting everything setup and optimized on the GTX 980's.

 

[iwajabitw]

Added the code Tony posted to my 64_config file. Waiting for the first unit to finish to compare to the second one with the edited file. I did click read config files in Boinc to see if it changed anything in real time, could not tell.

Edit: First file finished in 26min with an estimated of 21min from the start. Second file estimate is 11min. So that edit for the config file really shaves the time off.

 

[StefanR5R]

About the CPU tasks which Collatz Conjecture also offers:

I made some quick tests with E5-2696v4 (Broadwell-EP). This runs at 2.8 GHz which is its non-AVX all-core turbo.

Session 1, Windows:

• HT on, 50 % load: 5d05h average est. time remaining, Collatz Sieve, version not noted

Session 2, Linux:

• HT on, 50 % load: 2d20h average est. time remaining, Collatz Sieve, version not noted

Session 3, from there rebooted into session 4 with identical set of tasks, Linux:

• HT on, 100 % load: 4d23h average est. time remaining, Collatz Sieve 1.21
• HT off, 100 % load: 3d03h average est. time remaining, Collatz Sieve 1.21

Session 5, Linux:

• HT off, 100 % load: >30 days estimated time remaining, Collatz Sieve 1.20

My conclusion:

1. At least on Broadwell-EP + Linux, having Hyperthreading on and using all hardware threads gives better throughput than having HT off.
2. You cannot compare machine vs. machine unless you make sure that you loaded them with same tasks. (Or you wait until they finished their work and compare then on a PPD basis.)
3. The latter point means that my observation of much longer estimated runtimes on Windows were probably caused by different types of tasks loaded to the Windows machine. Less likely, but not impossible, is that the Linux application is much better optimized than the Windows port (I doubt that).

 

[GLeeM]

If you get a message saying: "This project is using an old URL."

And you are about to remove and add the new project URL: remember to save/copy your *.config file because it will be deleted when you remove the project!

Now I am back to 14 minutes per WU instead of 30

 

[iwajabitw]

I have the project running stock on 4 machines, no setup issues on Win7 or 10.

 

[TennesseeTony]

I have had some trouble getting multiple cards to run, alongside certain CPU projects. Collatz by default asks for 0.952 CPU's, which is a problem in certain instances. I have created an app_config.xml that drops the CPU requirement to 0.001, if anyone needs to use it.

    <app_config>
    <app>
    <name>collatz_sieve</name>
    <gpu_versions>
    <gpu_usage>1</gpu_usage>
    <cpu_usage>0.001</cpu_usage>
    </gpu_versions>
    </app>
    </app_config>

 

[StefanR5R]

About the CPU tasks which Collatz Conjecture also offers:

I made some quick tests with E5-2696v4 (Broadwell-EP). This runs at 2.8 GHz which is its non-AVX all-core turbo.[...]

• HT on, 100 % load: 4d23h average est. time remaining, Collatz Sieve 1.21

Guessing that these were tasks of the type with 36 k points/task, this means

• 7.3 k PPD per thread,
• 640 k PPD per dual E5-2696v4 box

which is at the order of 1/10 of a single GTX1080 with proper config.

@Markfw's Ryzen 1800X and 1700X apparently give 13.3 k PPD per thread, which is greater than just their clock speed advantage.