Custom Clients
log in

Advanced search

Message boards : News : Custom Clients

1 · 2 · 3 · Next
Author Message
Profile Kevin
Volunteer moderator
Project administrator
Project developer
Project tester
Volunteer developer
Volunteer tester
Project scientist
Avatar
Send message
Joined: 27 Jul 12
Posts: 507
Credit: 14,344,251
RAC: 1,727
Message 2139 - Posted: 9 Apr 2014, 9:00:17 UTC

Please don't use the custom clients - they don't always get the right answer.

I hope to build a release optimised clients after easter and we'll start testing the GPU version soon.
____________
Regards
Kevin
-----
International Centre for Radio Astronomy Research

Justberry
Send message
Joined: 26 Mar 14
Posts: 2
Credit: 481,591
RAC: 0
Message 2140 - Posted: 9 Apr 2014, 9:07:58 UTC
Last modified: 9 Apr 2014, 9:13:05 UTC

Can these errors be filterd out? because what we can read of tests, that it shows a huge increase in speed (350% to almost 6 times faster).
4 more wu's for 1. Given that 5% gives an error, Its still a huge increase

Freeze
Send message
Joined: 31 Mar 14
Posts: 2
Credit: 256,924
RAC: 0
Message 2141 - Posted: 9 Apr 2014, 9:15:22 UTC

I'm not sure if wrong results all error out. If any stay in, the results of the project are quite screwed. Say that a modified client gives the wrong answer continuously, and it gets verified by another modified client (who will insist the answer is right). Then what? It's a risk I don't think Kevin will take.

Maybe it's an option for the speed devils among us to tell what they used, so Kevin (and others) can profit.
I'd expect the standard things (-O3, -msse2, etc) are used already.

ps. do NOT use -ffast-math. Just don't...

Thedon46
Send message
Joined: 24 Mar 14
Posts: 1
Credit: 1,414,142
RAC: 0
Message 2142 - Posted: 9 Apr 2014, 10:19:27 UTC

What will be the consequence for those who continue to use it?
And more important, will the results obtained with the modified client count for the challenges such as http://www.theskynet.org/challenges/27 where a reward in dodge coins (real money) is involved?

Imho, you taking a very clear standpoint in these issues helps maintain a friendly community where people feel appreciated, satisfied and does not enforce unfair competition.

Profile Phil1966
Send message
Joined: 26 Jul 13
Posts: 7
Credit: 2,563,072
RAC: 165
Message 2143 - Posted: 9 Apr 2014, 10:34:43 UTC - in response to Message 2139.

Dear Kevin,

Are you planning to test/release CUDA as well as OpenCL WU's ?

Or one of them "only" ?

Thank You

Kind Regards

Philippe

bassiexp
Send message
Joined: 1 Apr 14
Posts: 3
Credit: 1,933,636
RAC: 0
Message 2144 - Posted: 9 Apr 2014, 11:57:26 UTC
Last modified: 9 Apr 2014, 11:58:45 UTC

Kevin,
Is there a way that those WorkUnits which have been processed with the optimized program.
So if it turns out these WU's aren't processed correctly you can resend them to be triple checked if there's really no computational errors.
Would be a shame if it makes part's of the project invalid.
This is also the reason why most of DPC stopt experimenting/using this program (Or never even started).

[DPC]_Fatal_Error_Group~Bubbles
Send message
Joined: 30 Mar 14
Posts: 1
Credit: 1,871,147
RAC: 0
Message 2145 - Posted: 9 Apr 2014, 12:03:09 UTC - in response to Message 2144.

Is there a way that those WorkUnits which have been processed with the optimized program can be recognized as such.
Fixed that for you. ;)

bassiexp
Send message
Joined: 1 Apr 14
Posts: 3
Credit: 1,933,636
RAC: 0
Message 2146 - Posted: 9 Apr 2014, 13:40:40 UTC - in response to Message 2145.

Oops, That is what I meant ;)

Grandpa
Send message
Joined: 26 Jan 14
Posts: 5
Credit: 68,242,750
RAC: 0
Message 2147 - Posted: 9 Apr 2014, 14:48:42 UTC - in response to Message 2139.

Please don't use the custom clients - they don't always get the right answer.

I hope to build a release optimised clients after easter and we'll start testing the GPU version soon.


Hi Kevin

I am currently using the custom client which I will not be able to shut down until I get home this evening, But the latest version has a very low error rate <1% which is lower than the stock client on my machines, are you sure it is the client causing the errors or is it OCed hardware causing the unusable results. My guess is it is a hardware problem rather than a client problem.

You may be noticing it more because of the shear volume of work being processed now, All of my Intel 4P and AMD 4P rigs are OCed and do have the potential to throw errors as are many other donor computers. If the skewed results are results coming from Hardware issues which I would tend to believe then stopping the modified client is not going to help anything especially since I am currently having a lower error rate with the modified client than I do with the stock client just a much larger volume of work is being done per machine I am running a little over 400 cores that have gone from 1 WU every 2 1/2 hrs to to 3 WU per hr. so a x7 increase in work produced in a day close to 30,000 WU's a day vs 4000 WU's a day so a person would be seeing close to a x7 increase in errors just from the volume of work.

I did not write it, I do know it was done with the knowledge of SkyNet and by the way the version I am using on the G34's it is using sse2 I am not sure what the Intel is using.

Hopefully you will respond before I get home from work this evening.

Thanks
Grandpa

Profile Coleslaw
Avatar
Send message
Joined: 12 Aug 12
Posts: 52
Credit: 2,702,356
RAC: 1,958
Message 2148 - Posted: 9 Apr 2014, 17:37:14 UTC
Last modified: 9 Apr 2014, 17:41:47 UTC

Yes, before the thread turns in to those people, here is the thread discussing it a while ago. http://pogs.theskynet.org/pogs/forum_thread.php?id=335

Maybe it's an option for the speed devils among us to tell what they used, so Kevin (and others) can profit.
I'd expect the standard things (-O3, -msse2, etc) are used already.

ps. do NOT use -ffast-math. Just don't...


Freeze, Kevin has already been given details of the optimized app we use. We have also posted links to it from our forums. It has been made no secret.
http://hardforum.com/showthread.php?t=1620596
http://hardforum.com/showthread.php?t=1801908
____________

Freeze
Send message
Joined: 31 Mar 14
Posts: 2
Credit: 256,924
RAC: 0
Message 2149 - Posted: 9 Apr 2014, 18:12:43 UTC

@Coleslaw: I'm trying not to derail, and not to blame anyone (after all, faster apps are a damn good idea :) ). I've missed the code NickOfTime posted, thanks for pointing me there.

All I wanted to make sure was that code produced is 100% safe, and not producing funny rounding errors. Those can be pretty much killing the whole project, which nobody wants.

Profile Coleslaw
Avatar
Send message
Joined: 12 Aug 12
Posts: 52
Credit: 2,702,356
RAC: 1,958
Message 2150 - Posted: 9 Apr 2014, 18:14:46 UTC - in response to Message 2149.

Freeze, I agree. And these concerns are exactly why I asked NickOfTime to contact Kevin from the very beginning. ;)
____________

Grandpa
Send message
Joined: 26 Jan 14
Posts: 5
Credit: 68,242,750
RAC: 0
Message 2154 - Posted: 10 Apr 2014, 2:47:26 UTC

OK all of the machines have been set to No New Work until I get a chance to redo them, hopefully this weekend I will have time, sorry if I caused any inconvenience

Empie
Send message
Joined: 24 Mar 14
Posts: 3
Credit: 1,007,652
RAC: 0
Message 2175 - Posted: 11 Apr 2014, 14:18:37 UTC
Last modified: 11 Apr 2014, 14:19:27 UTC

I'm having at least 4 workunits which are all marked bad. All of them are also cruched with the modified clients. From Grandpa we know he uses/used it, so I asked the other cruncher and he confirmed using it also.
First I thought something was wrong on my side, but now I highly doubt it.

It also looks like I'm seeing an increase in inconclusive results, which then get validated (or ended up as one of my 4 'bad' ones). It's not much, but still they are there.
Are more people starting to use it since word came out? If so, this isn't good for the reliability of the project.

Link:
http://pogs.theskynet.org/pogs/results.php?userid=15836&offset=0&show_names=0&state=5&appid=

Host 31689 is considered by me to be stable. Host 31671 is a fairly new machine, but up till now didn't give me any errors in BOINC.
I'm using the stock app on a non-OC PC with stock BOINC.

Btw, I'm not blamimg anyone for using a modified app or pointing fingers. Hope this gets fixed really soon.

Profile Coleslaw
Avatar
Send message
Joined: 12 Aug 12
Posts: 52
Credit: 2,702,356
RAC: 1,958
Message 2176 - Posted: 11 Apr 2014, 16:08:05 UTC - in response to Message 2175.
Last modified: 11 Apr 2014, 16:10:54 UTC

Empie, others have for sure used the optimized client. I have read at least 2 other teams other than my own who have mentioned it for sure. Most if not all of my team are pulling it off of their systems until Kevin reports back. We have also pulled the links to prevent any further downloads until we hear other news. I know many at DPC have also stopped using it. However, I wont confirm that all who are running modified clients have listened to Kevin's request. This is shameful because this says you don't care about the project. If the project says no, then that should be the end of it. It doesn't matter if you think you are right or if your app works better. We are here for the project and we play by the projects rules. As far as your results go, I don't know enough about the results to elaborate. That will be up to Kevin.
____________

Empie
Send message
Joined: 24 Mar 14
Posts: 3
Credit: 1,007,652
RAC: 0
Message 2177 - Posted: 11 Apr 2014, 16:44:56 UTC

Coleslaw, I agree with you.

Grandpa
Send message
Joined: 26 Jan 14
Posts: 5
Credit: 68,242,750
RAC: 0
Message 2178 - Posted: 11 Apr 2014, 16:46:46 UTC - in response to Message 2176.

I have found 2 instances where results using the modified client were confirmed as valid and results using the original client were confirmed a not valid. That is not good 1 is too many that tells me something is wrong either with the custom client or the original. I checked the stats today and the majority of the top 10 producers are still running a custom client it is pretty easy to tell they still have work in progress and their completion time is very fast compared to the stock client. http://stats.free-dc.org/stats.php?page=proj&proj=pog

Those that are still running a custom client need to stop running it. There is a very real chance you are doing damage to the project there is now evidence that I have found that points in that direction it is possible that bad results are being marked as good and good results are being marked as bad.

Yes I ran the custom client but I did not do it out of greed I did not think it all the way through and only looked at the error rate and thought I was speeding up the process, I did not even consider I might be messing up the science and that was my fault. Those that are continuing to run custom client days after we were asked to stop are only doing it out of greed. I personally believe there computers should be black listed and points be zeroed out. If mine get zeroed out that is fine also, I should have thought it through better.

I say shame on those of you that are still running custom clients after being asked to stop by the project my guess would be most if not all of you are aware of the potential problem and could care less.

DigiK-oz
Send message
Joined: 20 Aug 13
Posts: 5
Credit: 1,937,681
RAC: 0
Message 2179 - Posted: 11 Apr 2014, 17:39:04 UTC

Hold your horses. I agree, anyone still using the unofficial client should think it over, again...and again.

But, Some people are running several systems, which may not be available to them at all times. So, it might take some time to revert to the official client. Also, DPC has their annual stampede going on, so there will be users getting a very high score with the official client.

As far as I can see, most users in that top-10 have had a large increase in daily credit for a few days, but numbers are going down. Remember, due to validation of their WU's, the daily points will take a while to decrease. There are a bunch of users in that top-10 I know for a fact have either reverted to the official client, or never used the optimized one at all.

So, let's wait for a few days before naming and shaming, please.

Don't get me wrong, I fully agree that running an optimized client is simply NOT DONE if the project says no.

Profile Coleslaw
Avatar
Send message
Joined: 12 Aug 12
Posts: 52
Credit: 2,702,356
RAC: 1,958
Message 2180 - Posted: 11 Apr 2014, 17:49:01 UTC - in response to Message 2179.
Last modified: 11 Apr 2014, 17:49:55 UTC

Hold your horses. I agree, anyone still using the unofficial client should think it over, again...and again.

But, Some people are running several systems, which may not be available to them at all times. So, it might take some time to revert to the official client. Also, DPC has their annual stampede going on, so there will be users getting a very high score with the official client.

As far as I can see, most users in that top-10 have had a large increase in daily credit for a few days, but numbers are going down. Remember, due to validation of their WU's, the daily points will take a while to decrease. There are a bunch of users in that top-10 I know for a fact have either reverted to the official client, or never used the optimized one at all.

So, let's wait for a few days before naming and shaming, please.

Don't get me wrong, I fully agree that running an optimized client is simply NOT DONE if the project says no.


I concur with this. My shaming is pointed only to those who are knowingly continuing. I too took a day or two to finish the full removal even though most systems weren't actually using it. And I have read some users openly claiming to still be using it. I'm not pointing at any single team.
____________

Empie
Send message
Joined: 24 Mar 14
Posts: 3
Credit: 1,007,652
RAC: 0
Message 2181 - Posted: 11 Apr 2014, 18:07:42 UTC
Last modified: 11 Apr 2014, 18:08:26 UTC

DigiK-oz, was thinking of that too.
Also, if you take away points, that might be just as harmfull, maybe even more. You would then punish people that have had good intentions, like you Grandpa.
Not everybody is as active on forums and might not have read this topic, or has immediate access to the hardware they are running.

Either way, I'd rather opt for giving the people running the official client more points to compensate after a grace period since the start of this topic and resending the doubtful results in a different format when the new client arrives and let the old queue simply dry out.
But a simple "we fixed it" would suffice for me too.

1 · 2 · 3 · Next

Message boards : News : Custom Clients


Main page · Your account · Message boards


Copyright © 2017 The International Centre for Radio Astronomy Research