hostid=5393593 has reverted back to Stock since the list was published.

Claggy

No evidence of participation in these forums, but a member of a non-english-language team. I hope that's a good sign that the message might spread outwards through other means.

Edit - yes, they are discussing this problem on the SETI.Germany forums. Thank you.

Sten-Arne wrote:

But number one though, is to find an automated way to find and indentify those hosts that produces false results into the database. There's no way in h*** that it's going to be possible for volunteers to manually search through the system to find those hosts. It has to be done in some automatic database search fashion.

4202271
5257703
5293938
5354400
5393593
5396192
5445713
5467571
5472266
5478875

                                                              Joe

Yes, you have found a few of them. Was that made through manually searching the system the way we all can, or do you have access to the database through other means?

By manually searching, mostly last August. By August 25 I was no longer finding new host IDs by checking wingmates confirming Fermi false overflows so I judged the list of 22 hosts was complete for all practical purposes. Rechecking that list yesterday found 8 which were still actively using V12 on Fermi GPUs, and without being thorough I also picked up 2 additional IDs.

Whether 10 or 20 is "few" depends on point of view, and I didn't keep track of how many Fermi GPUs each host has. The false overflows only take about 20 seconds, and those VHAR WUs which the combination crunches correctly not much more time. So those GPUs will be wanting more than the base quota of 800 tasks per day. The frequent invalid cases will ensure that the quota doesn't often go much above that base, the fact that the false overflows are "Success" as far as the report goes ensures the quota will not stay below that base. So the quota system is at least helping somewhat to limit the problem, but it would be better if there were some way it could distinguish between a continuing problem and a temporary accident.

================================

Hosts which are running the correct software but still produce many false overflows is essentially an insoluble problem. BOINC advertises that simply installing the infrastructure and selecting projects to run is all that's required, and the majority of users essentially do so. Given that the crunching is done with consumer grade hardware, the projects must be prepared to deal with anomalous results. While those of us who regularly read and contribute to this forum would prefer it otherwise, some marginal hosts will continue to produce results which have similar effects to additional RFI. Luckily, CUDA false positives when running the correct software very seldom match so mainly impact efficiency.

                                                             Joe

I chased one of my inconclusives down to a wingman running a 295. half of it was turning in -9s the other was turning in good work. He may never know something is wrong if he just looks at his credits still rising and his RAC still holding.

---

PROUD MEMBER OF Team Starfire World BOINC

heres another guy with invald problems. Seems he gets 1 WU correct for every 5 he messes up. 5444296

In fact of his current valid WU's he does have invalid work getting credit[/url]

---

In a rich man's house there is no place to spit but his face.
Diogenes Of Sinope

Yep, and here's why

V12 modification by Raistmer

We need to really start harping on the fact this old app won't work with the new Fermi cards!!

---

PROUD MEMBER OF Team Starfire World BOINC

heres another guy with invald problems. Seems he gets 1 WU correct for every 5 he messes up. 5444296

In fact of his current valid WU's he does have invalid work getting credit[/url]

As someone else mentioned, how can he have a max tasks per day of 101 if he only gets 1 OK tasks for every 5 bad ones?

SETI@home Enhanced (anonymous platform, nvidia GPU)
Number of tasks completed 1657
Max tasks per day 101
Number of tasks today 585
Consecutive valid tasks 0

Another fermi with wrong opp app 5701024

Another fermi with wrong opp app 5701024

Seems he gets 1 WU correct for every 5 he messes up. 5444296

Both of those are members of team SETI.Germany, like the one who cleaned up his host earlier in this thread. I wonder whether we might have a distribution pattern here?

Another fermi with wrong opp app 5701024

Seems he gets 1 WU correct for every 5 he messes up. 5444296

Both of those are members of team SETI.Germany, like the one who cleaned up his host earlier in this thread. I wonder whether we might have a distribution pattern here?

I did some digging and his wingman is turning in invalid and errors. Plus the worst is for machine 5701024 a lot of the work looks like it validated for 4.0 credits. His wingman is also running GPU.

---

[/quote]

Old James

I am obviously missing something here, but:

Number of tasks completed 1663
Max tasks per day 100
Number of tasks today 585
Consecutive valid tasks 0

Really doesn't make any sense.

Bernie

---

As of now, this is as close as I can come to a complete list of hosts running the V12 CUDA application on Fermi GPUs:

3099502, 4202271, 5029073, 5149058
5231715, 5257703, 5293938, 5305178
5354400, 5357884, 5396192, 5444296
5445713, 5448502, 5467571, 5472266
5478875, 5541031, 5564376, 5701024

Not all of those hosts have results showing which were reported within the last few hours, so some could possibly have been fixed recently.

There are 24 Fermi GPUs in those 20 hosts.

                                                               Joe

I am obviously missing something here, but:

Number of tasks completed 1663
Max tasks per day 100
Number of tasks today 585
Consecutive valid tasks 0

Really doesn't make any sense.

Bernie

Number of tasks completed is a count of valid tasks done with the indicated application since the new credit system was started. "Number of tasks which have been granted credit" would be more accurate.

Max tasks per day is an accurate label only for a single core CPU. It's multiplied by 8 for GPUs at this project, and of course by the number of compute resources available for the application. The 100 value can be exceeded if Consecutive valid grows, and can be reduced by errors and invalid work. However, when it's below 100 each result reported by the host as a success tries to double it; any temporary value of 50 or more jumps to 100. If the application is throwing false overflows they are reported as successes just like a real overflow caused by the data.

                                                               Joe

Just glad that the PC's causing the issues aren't mine or any members of my team. In looking at the resources of the machines it is too bad that they aren't being well used. Has anyone PM'd the users in question? I believe that there is a proper optimization guide that was published by one of my team members - AndrewKing01 floating around here somewhere. I will dig it up if it can't be found.

The members of GPU Users group are always available to lend a hand to any user that is experiencing problems getting their machine to run as they should.

Todd
Team GPU Users Group Admin - World RAC Leader

---

Do you think it would do any good to post on the seti home paige. Just a short blurb that some cudas are running the wrong apps and all users should check their cards?

Or maybe a mass E-mailing to all users.

Or maybe we are worried over nothing, Be interesting what the labs take on this is.

---

[/quote]

Old James

Just glad that the PC's causing the issues aren't mine or any members of my team. In looking at the resources of the machines it is too bad that they aren't being well used. Has anyone PM'd the users in question?

I've been trying to PM the ones I catch as my wingmen but I'm not sure it will do much good. If they haven't paid attention to the forums they probably have email notification turned off.

---

PROUD MEMBER OF Team Starfire World BOINC

Do you think it would do any good to post on the seti home paige. Just a short blurb that some cudas are running the wrong apps and all users should check their cards?

Or maybe a mass E-mailing to all users.

Or maybe we are worried over nothing, Be interesting what the labs take on this is.

Nope since the apps are not technically setis and because these are obviously people that set it and forget it. they dont check in here, they don't post here, they don't monitor their machines after all the time people are told that they need to monitor their machines and check in to be sure some update hasn't occurred.

---

In a rich man's house there is no place to spit but his face.
Diogenes Of Sinope

As of now, this is as close as I can come to a complete list of hosts running the V12 CUDA application on Fermi GPUs:

3099502, 4202271, 5029073, 5149058
5231715, 5257703, 5293938, 5305178
5354400, 5357884, 5396192, 5444296
5445713, 5448502, 5467571, 5472266
5478875, 5541031, 5564376, 5701024

Not all of those hosts have results showing which were reported within the last few hours, so some could possibly have been fixed recently.

There are 24 Fermi GPUs in those 20 hosts.

                                                               Joe

One less now, 5541031 has updated.

                                                                Joe

I´m wondering how this can happen.

http://setiathome.berkeley.edu/workunit.php?wuid=671914757

I can´t find the credit multiplier in the fermi app.

---

With each crime and every kindness we birth our future.

I´m wondering how this can happen.

http://setiathome.berkeley.edu/workunit.php?wuid=671914757

I can´t find the credit multiplier in the fermi app.

The new credit system doesn't use anything from the app, the elapsed time reported by the BOINC core client (or CPU time from older versions of BOINC) is the basis of credits. That's combined with server averaging which produces what amounts to a benchmark for each application on the host.

The WU you linked has of course been purged from the database, with only about 6 hour purge delay it's best to quote pertinent details here.

                                                               Joe

From my latest run looking for anonymous platform hosts using outdated CUDA apps on Fermi cards, 5029073 has been updated and is no longer in the list:

Running Lunatics V12... 3099502 4202271 5149058 5231715 5257703 5293938 5305178 5346869 5354400 5357884 5396192 5444296 5445713 5448502 5467571 5472266 5478875 5564376 5701024
Running Crunch3r's 64 bit Linux r12... 5618971

                                                               Joe