Numberfields@home

Gilthanis

Gilthanis

[H]ard|DCer of the Year - 2014
Joined
Jan 29, 2006
Messages
8,730
I've been meaning to follow up on this thread.

First, to summarize: the 64bit linux app is about 4 times faster than the corresponding 32bit windows app. The 64bit linux app is twice as fast as the 32bit linux app. These times hold for the longer running cases (>5 hours). For cases that are less than 30 minutes, there is little improvement in the time. I believe the longer running cases are spending much more time in the deeper recesses of the algorithm where it's factoring very large integers, and this is where the 64bit version will outshine the 32bit version (I will need to use a profiler to prove this).

So it's apparent that a 64 bit version of windows would be helpful, at least for the longer running cases. To this end I've spent the last 2 weekends modifying the Pari library so that it would build with 64bit mingw. I now have a 64bit windows app that is giving the correct answers. I am currently running some timing tests. If these tests show the 2x improvement that I am hoping for, then I will promote this new app to the project.
Click to expand...

http://numberfields.asu.edu/NumberFields/forum_thread.php?id=132#775



IMPORTANT!

https://numberfields.asu.edu/NumberFields/forum_thread.php?id=355#2133

If you want your stats exported, you will need to check the consent box on the project preferences page.

In a couple days, the stats export mechanism will be changed, and if this consent is not given, then the default will be to NOT export your stats.

Sorry for the inconvenience, but this was necessary due to the recent GDPR regulations.
1. Go to your account page.
2. Click on NumberFields@home preferences.
3. Click edit preferences in the topmost section.
4. The consent box is the 4th one down.
5. Click update preferences.
 
Last edited:
Hmmm. I think a couple boxes are about to get Linux installed on them! That is unless running a 64bit Linux VM on a Server 2012 R2 Host would suffice? If so... that would be much easier. :p
 
It all depends on the gain versus the overhead of the VM... Being that numberfields seems to have significant increase in speed, I would say even a 64bit Linux VM would be faster than Windows on bare metal until they get the 64bit Windows app available.
 
For testing this I'm using two identical quad-socket E5-4xxx systems:

System 1 was wiped out and has a fresh copy Ubuntu 13 installed

System 2 has Server 2012 left alone, but added Oracle VBox. Installed Ubuntu 13 on here as well

Both Ubuntu installations were updated, then installed BOINC via their package manager.

It's already not looking so hot for the VM route. Despite being told to use 32 CPUs at 100%, Windows reports 65-70% CPU usage whereas the VM OS thinks it's using ~95%. On top of that, the GUI is soooo slow it's absolutely painful. Makes me think I should've used old 12.x instead.

We'll see how things add up in the morning. If it doesn't compare all that favorably, I'll wipe System #2 and put a fresh copy on it.
 
I just added 64 bit windows versions of the apps; your clients should eventually pick up the change if they haven't already. I've been testing them for the last couple days on my machine, so they should be good, but I'll keep an eye on them just to be safe.

These versions provide some speed improvement over their 32 bit counterparts, especially with the longer running WUs (those over ~1 hour). In some of the longer cases, the 64 bit version is almost twice as fast.
Click to expand...

http://numberfields.asu.edu/NumberFields/forum_thread.php?id=235#1284
 
As has been discussed in the message boards, there are a relatively small set of really long running WUs. Some of these may take up to 8 days or even longer.

As an incentive to users to not abort these WUs, double credits will be granted for them.

The bad work units have names of the form Qsqrt421_DS1x5_CV2_S815_*to* OR Qsqrt421_DS1x8_CV1_S815_*to*.

There are about 500 such WUs, of which ~150 are of the really slow variety.
Click to expand...

http://numberfields.asu.edu/NumberFields/forum_thread.php?id=263#1432
 
No, but it's something I'd like to do.

Several years ago, a volunteer wrote a GPU version that gave the correct results. The problem was that it still used the CPU for a big portion of the calculation and didn't improve timing very much. To make it a true GPU app, the PARI code (at least the functions we use) would need to be rewritten for the GPU. So it will be a non-trivial task. If you know any student looking for a class project, feel free to send them my way.
GPU Computing

If anyone knows someone interested
Click to expand...
Click to expand...
 
IMPORTANT!

https://numberfields.asu.edu/NumberFields/forum_thread.php?id=355#2133

If you want your stats exported, you will need to check the consent box on the project preferences page.

In a couple days, the stats export mechanism will be changed, and if this consent is not given, then the default will be to NOT export your stats.

Sorry for the inconvenience, but this was necessary due to the recent GDPR regulations.
Click to expand...

1. Go to your account page.
2. Click on NumberFields@home preferences.
3. Click edit preferences in the topmost section.
4. The consent box is the 4th one down.
5. Click update preferences.
 
After discovering the broken point system and doing some testing, Eric had asked me about a few things. He was going to check with the BOINC developers on the matter. He just gave me an update to his inquiry.

I spoke with Richard Haselgrove about it. I had been using the runtime credit option in the validator. Now that I am using CreditNew many of the problems should be fixed. CreditNew is a must when GPU apps are present, since the other validator options are susceptible to major abuse. However, CreditNew still has problems of it's own. Richard knows about this and is trying very hard to get the developers to address it. I get the impression none of the developers want to deal with it, so Richard has an uphill battle ahead of him. I doubt anything will change in the near future.

Anyways, that's the best I can remember of the conversations. The main take away is the validator is no longer using wall time, which enabled the "cheating".
Click to expand...
 
NumberFields@home just released a faster CPU app yesterday.

I have made some improvements to the cpu code. I am seeing somewhere between 30% and 40% speedup, depending on the case.

I have deployed the new executable for 64bit linux. I am currently trying to get it to build for 64bit windows. I hope to have that ready by tomorrow sometime.

The next step is to use these same tricks on the GPU versions. GPUs can be very finicky, so this might take several weeks before I have something worthwhile.
Click to expand...
 
You must log in or register to reply here.
Back
Top