Hi all,

I have submitted some new work units that will replace some I submitted earlier in the week. The names will be "NATHAN_FAX3". These tasks are in the true spirit of the long queue, and will take about 12+ hours on the fastest cards. Some have already been returned and indeed have been around 13 hours. This is markedly longer than what you have expected traditionally, but we really want the long queue to be for critical tasks, computationally intensive tasks, and the like. I suggest you all take note of how these tasks run on your computers and be mindful of temps and errors as you start to receive them.

I have noticed some crunchers expressing concern/dismay that perhaps they will not be able to get the 24h bonus with such long tasks. We are mindful of that concern, and will keep an eye on this group as an experiment. If we think it is too unfair to people with fast but not the fastest cards, we'll be sure to correct that in future groups. But the less send/recieve we have to do, the better. We are also mindful of the fact that longer tasks might be more susceptible to errors/crashes, and we want to see how this goes. I'll be looking out for the severe error percentage over the next few days for any problems.

Also, a note about tasks beginning with NATHAN_FA... These tasks are unique in that they are quite large simulations, compared to many others we have done in the past which are smaller (bigger biomolecules mean bigger simulations). They not only take longer per step, but require more memory. Cards with lower memory (below 1GB) may suffer additional performance loss. There is nothing we can do about this, unfortunately.

Happy crunching.

Nate

Thanks for the info.

This is markedly longer than what you have expected traditionally, but we really want the long queue to be for critical tasks, computationally intensive tasks, and the like.

but we really want the long queue to be for critical tasks

I am expecting to run them in 32h aprox. I want to know in these cases that take more than 24h if you prefer them to be processed or that I switch off the "long runs" option.

---

HOW TO - Full installation Ubuntu 11.10

Regarding the 24 hour limit, I also am concerned about this because I have been experiencing a problem with the connection on upload which is apparently unrelated to either my system or the servers for GPUGRID.

When uploading a completed work unit, the upload invariably fails on one file. If there are successive failures, BOINC will retry, and BOINC adds time to the retry delay each time so that in the worst case, BOINC might not retry for up to 8 hours or more. There is, apparently, nothing that I or GPUGRID can do about this, and if the worst should happen, then it appears that my machine has not returned the work unit within the 24 hour limit - when it has, in fact, completed the work unit and made best attempts to return it.

Personally, I think the 24 hour limit is completely unfair, especially in cases like this where my machine and, indeed, I, have done everything humanly possible to return the work unit.

I understand what you are trying to accomplish by the 24 hour limit, however, I do not agree with the implementation.

---

Man, I thought I was going crazy when I saw this task running for over 15 hours, before I read this post. I thought my card had downclocked or something, but windows wasn't reporting anything unusual. I just let it run and it finished fine, in about 20 hours on my overclocked GTX570.

I did receive consistent credit with other "NATHAN_FAX" tasks -- the whopper was 114,000 in credit.

If I had a less of a card, the 24 hour bonus would have been in jeopardy. I would prefer if the tasks ran in the 8-12 hour window, but one of these occasionally doesn't matter (as long as it doesn't fail after 95%).

They not only take longer per step, but require more memory. Cards with lower memory (below 1GB) may suffer additional performance loss. There is nothing we can do about this, unfortunately.

I have 3034Mb (3Gb) on each of my two GTX 580 cards. (Most of which usually appears unused for GPUGrid tasks).
Is it possible for GPUGrid to allocate the tasks with high memory requirements to such computers?
If not, why not?
It is a high-capacity resource (6Gb) going to waste, as far as I can determine.

If you look at task properties, from Boinc, you will only see how much system memory is being used. You would need to use a tool such as GPUZ to see the amount of GPU memory being used on your card.

How much is being used? 1243MB of my GTX470's 1280MB is being used - which explains the higher temps. ~450MB is more common for GPUGrid tasks. I expect W7 may be further restricted by this, as the operating system grabs some GDDR for itself.

Typically the amount of memory being used is directly related to the number of shader blocks (cuda cores). So if 1243MB is used on a GTX470 (with 448 shaders), a GTX580 should scale to, 512/448X1243MB=1420MB. It normally scales down quite well too; so lesser GPU’s (with less shader blocks) don’t suffer from having less GDDR (GTX460 with 512MB is normally on power with the 1GB version).

The amount of GPU memory used is also experimentally confined by the number of molecules. However, if there was a need to use more memory, to run experiments with larger molecule count, GPUGrid could run one task across more than one GPU (for those with multi-GPU setups), &/or utilize the extra memory on cards such as the 3GB GTX580.

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

Man, I thought I was going crazy when I saw this task running for over 15 hours, before I read this post. I thought my card had downclocked or something, but windows wasn't reporting anything unusual. I just let it run and it finished fine, in about 20 hours on my overclocked GTX570.

I thought about aborting one myself. Luckily when I saw one it only had a hour left so I let it finish. These are running at around 20 hours on a stock clocked GTX470 in Linux. Unfortunately with a 115 GB upload I cannot get them turned in fast enough.

I have my cache set to 0.01 but they still download a few hours before they start so I run out of time.

I got one of them finshed too. I was only wondering why it gives only 30000 credits for 35000secs (FAX) instead the (CB1) nearly 25000sec for 35000 credits ^^

http://www.gpugrid.net/workunit.php?wuid=3247302

Now i got a second one. i will look how this one is going ^^



I think on a 560TI it would take ~12,5 Hours, the 285 took little bit under 12 Hours.

---

DSKAG Austria Research Team: http://www.research.dskag.at

Man, I thought I was going crazy when I saw this task running for over 15 hours, before I read this post. I thought my card had downclocked or something, but windows wasn't reporting anything unusual. I just let it run and it finished fine, in about 20 hours on my overclocked GTX570.

I thought about aborting one myself. Luckily when I saw one it only had a hour left so I let it finish. These are running at around 20 hours on a stock clocked GTX470 in Linux. Unfortunately with a 115 GB upload I cannot get them turned in fast enough.

I have my cache set to 0.01 but they still download a few hours before they start so I run out of time.

Set your cache in BOINC to 0 and connect to 0 and then it will complete before downloading another one.

---

Radio Caroline, the world's most famous offshore pirate radio station.
Great music since April 1964. Support Radio Caroline Team -
Radio Caroline

How much is being used? 1243MB of my GTX470's 1280MB is being used - which explains the higher temps. ~450MB is more common for GPUGrid tasks. I expect W7 may be further restricted by this, as the operating system grabs some GDDR for itself.

My current memory usage is (running NATAHAN_FAX3 task, and one NATHAN_CB1 task) ~700Mb=23% @71 Celcius &
~610Mb = 20% @65C.

Again, this seems like a curious waste of RAM.

That's around half what I expected (1420MB). So why is a GTX580 using half the RAM it should be? GDDR usage doesn't tend to change too much. So is it the operating system, driver, or app?

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

How much is being used? 1243MB of my GTX470's 1280MB is being used - which explains the higher temps. ~450MB is more common for GPUGrid tasks. I expect W7 may be further restricted by this, as the operating system grabs some GDDR for itself.

My current memory usage is (running NATAHAN_FAX3 task, and one NATHAN_CB1 task) ~700Mb=23% @71 Celcius &
~610Mb = 20% @65C.

Again, this seems like a curious waste of RAM.

Do you know if an app_info file will work here? If so you may be able to set it up to run two units at once on each card. That however WILL stress the limits of your cards AND crank up the heat output of your cards too! This CAN significantly shorten the life of your cards.

I got one of them finshed too. I was only wondering why it gives only 30000 credits for 35000secs (FAX) instead the (CB1) nearly 25000sec for 35000 credits ^^

http://www.gpugrid.net/workunit.php?wuid=3247302

Now i got a second one. i will look how this one is going ^^



I think on a 560TI it would take ~12,5 Hours, the 285 took little bit under 12 Hours.

Well that is significantly better than what my 560ti is doing.

currently running http://www.gpugrid.net/result.php?resultid=5098225

overclocked to 950/1900/2007@ 1.025v and it's showing 13:21:xx time used with 10:40:xx to go. So that is going to be just over the 24hrs, NOT including upload time. how big are the upload files on these anyway?

additional information for those interested. It's a 2gb model currently using 1240 mb VRAM and 263Mb system ram.
4.5Ghz i7-2600k

I got one of them finshed too. I was only wondering why it gives only 30000 credits for 35000secs (FAX) instead the (CB1) nearly 25000sec for 35000 credits ^^

http://www.gpugrid.net/workunit.php?wuid=3247302

Now i got a second one. i will look how this one is going ^^



I think on a 560TI it would take ~12,5 Hours, the 285 took little bit under 12 Hours.

Well that is significantly better than what my 560ti is doing.

currently running http://www.gpugrid.net/result.php?resultid=5098225

overclocked to 950/1900/2007@ 1.025v and it's showing 13:21:xx time used with 10:40:xx to go. So that is going to be just over the 24hrs, NOT including upload time. how big are the upload files on these anyway?

additional information for those interested. It's a 2gb model currently using 1240 mb VRAM and 263Mb system ram.
4.5Ghz i7-2600k

just as an additional to this, I am now showing that it will take approx 25 hours to complete this task. apart from missing the bonus points, 25 hours really is too long for a workunit, even a big one. I know this is only a 560ti, but a 2gb model running at 950 should be able to do one quicker.
I'll try running one more after this one, just to confirm the timeframe. If it is correct around 25 hours, well I guess I won't be doing them anymore :(

Simba: Finally, runned a real FAX3 unit now. 30-31Hours runtime (285) with additional 12h uploadtime ^^ So your not alone with the 560TI ^^ (25hours is not bad for this card!) I will not run any FAX3 WU on my 560TI cos it computes to long. I often run the GamingPC over Night for 2 or 3 WUs after i was gaming. But now i dont continue that, cos i could get a FAX3 while i´m in sleep and want to turn the PC off in morning, not waiting aditional 20hours or so ^^

We still get bonustime for report it under 48hours, but not as much as under 24hours.

---

DSKAG Austria Research Team: http://www.research.dskag.at

Sorry for the delayed response, I have been at a conference for the past two days. Considering the responses and the stats we have gotten, it definitely seems that these are a little too long. There are a lot of people who complete the tasks on time but then have trouble uploading them in time for the bonuses at both the 24 and 48 hr limits (116mb is a large upload size). More importantly, there also seems to be more severe errors than are typical, though it's not clear why that is right now.

Therefore, I am going to modify and rename these tasks. They will be renamed to NATHAN_FAX4 (surprise!), and they will run for ~66% of the time of the FAX3 jobs. My checks here indicate that they will run for 8.3 hours on GTX 580, so remember that they are still not like short tasks. People who are not using the GTX 500 or 400 series cards should consider crunching only on the short queue if you are concerned about runtime, bonuses, failures, overheating, etc. Ultimately, the choice is yours. There is still a 25% bonus for finishing before 48 hours.

Hopefully this hits the sweet spot we and you guys have been looking for. This is a learning moment for us and it has been good to get some data on the behavior of such long tasks.

Nate

Wiyosaya: Definitely an unfortunate situation. I'm not sure what we can do on our end, since we can't change how BOINC works. I don't think we can do anything, but will ask.

Michael Kingsford Gray: We can't choose which computers get which jobs. The reason is that the BOINC system is not set up to allow that, and we can't modify it to do that. We can make different queues, as we already have three (ACEMD2 aka short, ACEMDlong, and ACEMDbeta), but making too many queues makes it confusing for crunchers and also increasingly complicated for us. The only people that could fix it to a better system are the people who maintain the BOINC software.

dskagcommunity et al.: I should clear up confusion about the different "FAX" named tasks. The first round of tasks, titled "NATHAN_FAX-" (with no number after) were supposed to be long like the FAX3 tasks. However, there was a problem with our test software and it underestimated runtimes (but correctly calculated credit). Those have been replaced by the FAX3 tasks. There were "FAX2" tasks, but they were stopped before anyone recieved them because the credit was incorrect by a huge factor (a human mistake, I failed to edit a line in the submission to the server). As you know the "FAX3" tasks are the very long tasks, which have proven to be a bit too long for everyone's taste, so I will modify them to run for shorter times as I explained above. I will rename them to "FAX4" in order to make the distinction clear. Sorry for all the different task names, but I need to be sure I can keep track of the simulations, and what data comes from where.

Personaly i think 66% is a real fair alternative :)

---

DSKAG Austria Research Team: http://www.research.dskag.at

Hopefully this hits the sweet spot we and you guys have been looking for. This is a learning moment for us and it has been good to get some data on the behavior of such long tasks.

Thanks for this response from you and the project. :-) I just got one of the FAX4's and should be crunching it on my 570 in about 90 minutes.

---

I'm just starting on a FAX4 and while my gpu idled for few seconds I saw in GPU-Z that the memory used was 150mb for Win7.
Now with the card loaded with FAX4 it's at 1232mb so that's a 1082megabytes usage for FAX4. Memory controller load is at 28-29% with 800mhz memory clock (gpu-z reading)

One thing I do NOT like is the gpu load. With other long runs and other projects the gpu load is anything between 94-99%. Now it's at 86-88% ?!
CPU is 2500K@4500Mhz
I have swan_sync=0 and only three cpu tasks running along with gpugrid.
Do I really have to put my cpu at higher clocks or what?
This is the task http://www.gpugrid.net/workunit.php?wuid=3264669
E: Well I bumped the clocks to 4800Mhz and it really didn't affect the gpu load %, still around 87-88%