Hello: I would like to comment on a matter which does not understand.

I see that there are tasks that we have made TWO users, completed, validated and marked; i.e. we have doubled the work. Best regards

You'll find that the first WU is over 2 days old without a result so it is issued to a second host.

The reason they do this is because they require fast and reliable completion as the next WU in a job depends on the first one being completed and can't be sent out before.
____________
Radio Caroline, the world's most famous offshore pirate radio station.
Great music since April 1964. Support Radio Caroline Team -
Radio Caroline

Hello: From a few days ago the number tereas we performed in duplicate (I + a collaborator) and not by spending two days between sending each of them, in these last few hours have passed.

If there is a coordination problem on the server, it seems a waste of labor. Greetings.

Hello: So far I see this, there were two more but now indicate miscalculation ...?. Will report if I find others. Greetings.

http://www.gpugrid.net/workunit.php?wuid=2462615

Hello Again it happened that a task is sent to another user before the end of two days of processing in the first ...!

Proceed to cancel the task because it is losing work with it, but it is unfortunate the time I lost my ...! Greetings.

Work: http://www.gpugrid.net/workunit.php?wuid=2514023

Hi, again a task has been sent to two users on the same day (three hour time difference) I think is worth reporting if a server failure or any special circumstances of the task ...? Greetings.

Workunit 2521888

http://www.gpugrid.net/workunit.php?wuid=2521888

At the moment I suspect that a lot of work is being wasted due to this policy. Don't forget that BOINC is a multi-project environment and the scheduler takes into account the project resource share and the deadline when scheduling tasks. The problem is made worse in that BOINC can take up to a day to send the results back.

A much better solution would be to make the deadline 2 days and let BOINC work as it was designed. You will get more complaints about the short deadline but it's better to be open about it instead of wasting resources. The users can then decide for themselves whether GPU is for them.

It's only recently I've become aware of this policy and it's very frustrating to realise that probably every one of my tasks has been re-issued. What a waste!

Hmm I agree with making the deadline 2 days instead of 5. I often get sent tasks which are still running on slower GPUs, my WU will complete and it'll still be running on the other host, just a waste of resources if you ask me. Would also, hopefully, stop boinc downloading 2 GPUGRID tasks (my buffer is set to 1 day because I need enough einstein WUs to feed my GPU which runs 4 of those at once)

Yes, I usually don't run both GPU projects at the same time. Also use report tasks immediately. I myself have no trouble completing WUs in time.

Running 4 einstein WUs works great btw. If I'm not using the pc it completes them in about 95 mins (with 4 CPU tasks running alongside). Running 3 or 2 only makes them complete a bit faster. But not enough to cancel out running 1 or 2 tasks less. Annoyingly enough even 4 einstein WUs only get GPU usage up to a max of about 80%.

By all means reward users for a fast turnaround, but I don't think it's a good idea to penalise users from returning tasks within the deadline period.

Not sure I understand what the issues are around BOINC's lack of up-time, however the last thing you want is for users to stock up their buffer above 2 days.

For info I've never had any problems running multiple GPU projects at the same time. BOINC takes care of this rather well and ensures tasks are (almost) always completed within the deadline, which brings us back to the original problem.

What's causing the waste here is the fact that tasks are resent before the first computer misses the deadline. The concept is simple...decide how soon you need results back and set the deadline to that many days. If they want them back in 2 days then grow a pair, draw the line and set a 2 day deadline. If they feel 3 days is adequate then make the deadline 3 days.

Setting the deadline to 5 and then resending tasks after 2 days is just plain dumb because it's inevitably going to end up producing a lot of wasted (duplicated) effort. It's better for systems that can't return a task in 2 days to crunch somewhere else rather than waste their time duplicating work here.

I agree with Dagorath on the deadline.

Boinc calculates the ratio of time the computer is on and running Boinc. If Boinc thinks it is not on long enough to finish a task in 2 days then Boinc will not ask for a task.

2 days would be long enough for many people for any tasks, but not long enough for many people trying to run long WU's.

Boinc sees different tasks as all part of the same GPUGrid propject, so just trying to run one long task might throw it out and prevent someone that would be capable of running normal tasks from getting any.

Then there is different Boinc versions (some might behave slightly differently to others) and what would happen if you had a poor GPU and replaced it with a high end GPU?

Basically I think a bit more leeway is called for than a 2 day cutoff.

Ideally Boinc would better understand separate task types and their requirements and even allow the user to more exactly specify what type of task and when these tasks should run. So GPUGrid has to be a bit more flexible and try to facilitate more people and operate within Boinc's restrictions.

The research team also spent a lot of time and effort getting to where we are now. At some stage (now) they have to concentrate more on the Science and maintenance (new GPUs/drivers/CUDA versions) and less on the project setup and development. Also, changes always annoy someone.

I was explaining what might happen if we had a 2day deadline, speaking hypothetically, rather than explaining the existing 5day system. I agree that the system should be reviewed periodically, and the deadline probably reduced, but it's not my call and I don't have all the info.

Resends are always going to happen (tasks fail, as do GPU's, computers, routers).

Even if the deadline was 2 or 3 days, tasks would still fail/not be returned in time and have to be resent.

The number of resends might actually rise if the deadline was changed.

I don't have the figures so I can't work it out, but the team could do some stats to work out in advance if it would expedite the project overall, or hinder it.

If they were really up for the challenge they could write a program to continuously analyze returns and self-regulate the deadline within boundaries, but I expect Boinc would backfire.

At least long workunits should not be sent to slow hosts. They have no chance to return them even in 5 days. I'm still getting resent tasks, which are still being processed on GT 9500 or GT 9600 cards. I don't feel like warning these users one by one. I wrote about it in march.

Hi: Following the points made here, also estimated that if the project needs the results in two days makes no sense to confuse users with the story of five days and generate duplication.

It generates the five days that users with GPUs that need three or four days to complete a task (not learn it in two days) participate miserably losing time, money and enthusiasm to participate in these tasks.

A) Identifies that GPUs is installed, if it is not recommended simply not accept the participation in this project.

B) When the time came to execute the task, if it has not been completed, canceled automatically (Boinc can do it perfectly) and now if it is referred to another user.

When you cancel once or twice a task and you find out what the problem and project requirements in which you participate. Greetings.

We have increased the duplication time from 2 to 3 days.
gdf

Hi: The two tasks that I am running right now are in the process by another user.

One of them has seven tasks in the process (with NVIDIA Quadro 4000) lowered on day 6 so have already been forwarded to as many users ... !!!

I know not much, I say so, but I keep insisting that it makes sense to keep two dates to complete the tasks, the famous 3 to 5 days. Greetings.

Hello: Every time I understand forwarding policies least half-finished tasks ... Now I get one that makes 24 hours sent to another user ... as 3 days or case. Greetings.

Task: http://www.gpugrid.net/workunit.php?wuid=2579334

Thought I already explained that situation?

Anyway,
The 1st resend was after 3days - the task was not returned, this is normal.
The 2nd resend occurred after that because the task failed (user aborted in this case) - this is also normal; a failure also triggers a resend, even if a resend is already in progress.

The system certainly does need to be revised but let's keep it simple rather than make it more complicated. The suggestion to issue resends only to CC 2.0 hosts would require new code on the server side, I think.

The simple, easy to implement solution is to get rid of the double deadline that now exists. Having a 3 day deadline plus another 5 day deadline leads to waste as has been demonstrated numerous times in this thread. Have 1 deadline and 1 deadline only. If you think I'm asking to have the 3 day deadline increased then read the post again because that's NOT what I want.

BOINC server code already has an option to issue resends to fast, reliable hosts. The code to do that was tested and proven several years ago. Why not use that instead of writing new code to issue resends to CC 2.0 hosts?

Perhaps this system needs to be revised, but the importance of returning the task increases following a failure. As to why the user aborted, who knows. Usually they just think it will take too long, but they might have a problem with it, which would make it more important to return the task. The server cannot know that the first resend was/is running smoothly. Should one of the resends finish before the other starts (when someone has their cache too high) the other resend will not start (assuming intervening communications).

Suggested revision:
First resends only to be sent to CC2.0 cards
Second resends to be sent internally

Hello: Following what I promised, I put a link to a new task duplicate.

http://www.gpugrid.net/workunit.php?wuid=2589619

If the term to complete a task was bound and 3 days, this circumstance would not occur.

It is a partner with a GT9600 that takes at least 44 hours (nonstop) to execute a task is working with the idea and hope to run a useful task, when in reality it is not, all tasks have doubled. Greetings.

Hello: Again is coming more duplicate tasks

The last four tasks (including the two that I am performing at this time) are underway by other users that have exceeded initial three days.

Honestly is a topic that I will never tire of repeating and claiming for the loss of effort (and money) it means ... regrettable. Greetings.

Hello: We are the same, the two tasks I have just received are also duplicated.

This time one has been completed by someone else soon after I start the same task (one hour), have been canceled and proceeded to ask a new one (it is not published ... thank goodness).

I keep arguing, it seems that without success, that if the program needs the results in - three days - this is the real deadline to complete the task and nothing more ... is simple ... no? Greetings.

I have just removed the duplication mechanism as announced in another thread.
It will take few days before all duplicated workunits disappear.
Deadline for now stays at 5 days.

gdf

Hello. Thanks for the change.