Project News - AQUA@home

[Dataman]

AQUA has work again.

EDIT: They have a bug ... You need version 6.45.0 or 6.60.0 or higher.  biggrin Should be 6.4.0 You can't get work until they fix it.
EDIT: They have it fixed now. Seems "testing" has become a lost art. 

[veebee]

hmmm, methinks I might swap all my machines over to windows.... I have 1 machine (a q6600 at stock - WinXP 32 bit) which got over 330 cr/ hr on aqua, while my 64 bit linux machines only get around 50 - 55 cr'hr.

I REALLY wish I coul dget a working install disc of th RC Win 7 x64... I have burned half a dozen and NONE work.... WTF?

[BF]

hmmm, methinks I might swap all my machines over to windows.... I have 1 machine (a q6600 at stock - WinXP 32 bit) which got over 330 cr/ hr on aqua, while my 64 bit linux machines only get around 50 - 55 cr'hr.

I REALLY wish I coul dget a working install disc of th RC Win 7 x64... I have burned half a dozen and NONE work.... WTF?

PM me your address and I'll send you a working one.. 

[beakerulz]

Ive put some of the servers I have access to on this project and the results are amazing! the blades are dual socket 4 core, so yeah 8 cores to run on this. I have them all running Windows 2008 x64 and they are flying! Less than 24 hours and they have already clocked up 500k! man this is a good one for credit!

The one R900 that is a quad core quad socket, (16 CPU's) runs 8 cores against each of the 2 work units it downloads so it completes 2 workunits in about 7-8 hours. will do 6+ workunits a day. amazing stuff, i only just got onto this project but i think i should have awhile ago! Can only imagine with graphics card what it might do? If I had a good GPU card in this server would it make a big difference? Im thinking if it will to put one in but dont understand how one GPU card relates to 16 cores?

My servers are showing in the computers section of the Aqua@home if you want to see what they are.

Thanks everyone for your help and comments on what to do / setup options etc... i thought i needed an Nvidia card too or something, but the CPU stuff seems to work just fine!

Thanks

[kashi]

Actually the credit rate dropped over 50% when the most recent batch of 128 qubit tasks was released. It is still excellent but not as sensational as previously. Glad they released some more tasks so you could have some credit fun at AQUA, so you didn't miss the boat after all, just the really fast boat. :) That last really fast boat only ran for 2-3 days though and you can't cache more than 2 tasks, so everyone is now in the same boat.

I noticed your 3 GHz dual socket blades perform very well on AQUA. They complete a 128 qubit task in about 4 hours. My W3520 took 5.3 hours for the one 128 qubit task I have completed and received 6,270 credits, so about 148 Cr/Hr/Core.

Now that it is more generally known I can further explain that the reported Run time on the task results AQUA webpages is not related to BOINC Manager Elapsed time or CPU time in any way. It is the difference between the time you received the task and the time you reported it. If you held a task for a week before you processed and reported it then the "Run time" reported would be a week not the 4 hours you took to process it or your CPU time which would be approximately 24 hours at 75% average CPU utilisation on 8 cores. So the website reported "Run time" is not run time at all, you could call it AQUA server cycle time I suppose.

AQUA requires BOINC version 6.2.xx or higher.

I don't know the actual figures but from what I read the GPU application is not as efficient on AQUA as the multicore CPU application. Many with nVidia graphics cards run GPUGrid as their preferred GPU project, however some seem to have problems with GPU drivers. nVidia cards can run on more projects than ATI cards.

For excellent GPU credit though MilkyWay on a compatible ATI card is currently unmatched because the MilkyWay ATI application was developed to take advantage of the double precision performance of recent ATI cards. My ATI HD 4890 currently receives about 100K per day on MilkyWay. My highly overclocked HD 3850 received about 27K per day. It has been announced that the credit rate will be reduced but luckily for me the MilkyWay developers haven't got around to doing it yet.

Collatz Conjecture also looks promising for GPU cards both nVidia and ATI. The credit rate there on a HD 4890 is currently about 20% of MilkyWay ATI, but will probably receive a boost when Cluster Physik/Gipsel and Slicker release their new optimised application. Slicker has an unreliable internet connection so the server is usually down for part of every day.

[Dataman]

From BOINCAdmin at AQUA

"Hi all,

Considering the slow pace at which CUDA work units are being removed from the queue, we have decided to stop submitting CUDA work.

To elaborate more: We intended to use the CUDA version to solve 8-, 32-, and 48-qubit problems while the MT version was doing the rest. However, we are already at the 160-qubit level with the MT work units while the 32-qubit CUDA work units are still not finished.

The reason is a mixture of not having enough volunteers with CUDA cards and slower execution of the CUDA code compared to the MT application.

We ask our volunteers to please stay attached to the CUDA queue because we may submit jobs there if the need arises.

Thanks!"

*****

[WikiWill]

A forum post from Neo, project admin:

Some have noticed that the queue is empty now, and hopefully it won't be for too much longer, but as mentioned in a post in another thread, the next version of the app isn't compatible with the previous version. We don't want to risk crashing on everyone's computers, and although BOINC isn't very likely to download a new version of the app while crunching an old unit, it's happened before.

We also need a bit more time to do the new preprocessing phase for the units. Instead of doing it as part of each workunit as before, we're computing them ahead of time to make sure that the same values get used for each of the 50-100 copies of each problem instance that are sent out. It takes a bit of time to do these since we haven't yet BOINC-ified it enough to do a public release as a separate app. The simulation results we get from this new calculation look rock solid even for problems that were really though to simulate.

The workunits may end up being a bit longer per qubit with the new app, but we won't be simulating large problems (at least with the first batch), so they shouldn't be nearly as long as the current 200-8M's.

I'm thinking that early next week would be a good time to start the first run of the new app. Hopefully we'll get enough preprocessing done over the weekend. Thanks for your patience and all of the great help you've given us already! :)

So no new Aqua work for a few days.  Two of my machines have run out already so are back on WCG for now.

[kashi]

My WCG will be getting a little boost too. Just went gold for HCMD2 once pending is in and now working on Flu gold. :)

[beakerulz]

A forum post from Neo, project admin:

Some have noticed that the queue is empty now, and hopefully it won't be for too much longer, but as mentioned in a post in another thread, the next version of the app isn't compatible with the previous version. We don't want to risk crashing on everyone's computers, and although BOINC isn't very likely to download a new version of the app while crunching an old unit, it's happened before.

We also need a bit more time to do the new preprocessing phase for the units. Instead of doing it as part of each workunit as before, we're computing them ahead of time to make sure that the same values get used for each of the 50-100 copies of each problem instance that are sent out. It takes a bit of time to do these since we haven't yet BOINC-ified it enough to do a public release as a separate app. The simulation results we get from this new calculation look rock solid even for problems that were really though to simulate.

The workunits may end up being a bit longer per qubit with the new app, but we won't be simulating large problems (at least with the first batch), so they shouldn't be nearly as long as the current 200-8M's.

I'm thinking that early next week would be a good time to start the first run of the new app. Hopefully we'll get enough preprocessing done over the weekend. Thanks for your patience and all of the great help you've given us already! :)

So no new Aqua work for a few days.  Two of my machines have run out already so are back on WCG for now.

Thanks mate, made it real easy with you posting whats going on to figure out the status of the project! its is easier to look on the forum to see what you have posted then it is to try find on their website etc.. thanks for the heads up was wondering why my machines have run out of work... ok ill give them something else to do for a while!

[Robert]

 :penguin:
Thought I'd pop in to see if there was any update on AQUA work.
Nothing yet?

[BF]

There is a possibility of some 8 and 16 qubit units being released by Wednesday. Admin staff are working on both server problems and finalising the new app.  I might put 1 or 2 cores on when work is released, but keep the rest on the WCG AA.

[Dataman]

Yesterday from AQUA.

"My apologies once again for the lack of work, and while I'm running the AQUA test suite, I'd like to explain the ton of changes over the last few weeks that to lead to this craziness.




•Our Monte Carlo expert came up with a way of getting good simulation results on problems with small gaps, but it requires that a big preprocessing step for each problem to be simulated before we send out the 50-100 copies of each problem. We split the app into the two phases using a command line parameter, and had to change a few other parameters.

•The adjusted simulation gathers much more data than before as part of getting better results, making an unoptimized part of the app take about 90% of the time instead of 20% before, so I rewrote it to work completely differently and it should now be about 40%-50% of the time.

•We now have a new annealing schedule (how the weights of the "quantum part" and "classical part" change over time) that is precomputed to match what our next batch of hardware will be doing. Another few changes to command line parameters for this.

•The command line and program initialization were a mess at this point (and probably buggy), so I cut out huge chunks and restructured them a ton. It's looking almost not-embarrassing-to-publish now, but this introduced a couple of bugs that are now fixed.

•The first tests on the next batch of hardware will have input values with 2 bits of precision (BOP), whereas our previous simulations had 4 BOP. However, the 2 BOP problems have relatively huge (like 10x larger) energy gaps, and the values we're using to get the energy gaps are garbage for large gaps, so we got only garbage results. This can be partly solved by scaling down the energy by a factor f, then multiplying the gap sizes by f, but not all problems work for the same f. We'll probably just use a value that works for almost all problems (5 seemed good).

•In scaling down the energies simulated, we may need to scale down the simulation temperature, and the optimized data gathering code had a bug for when using a non-default temperature, which is now fixed.

•The exact gap results we were comparing against for the 2 BOP problems were found with the wrong annealing schedule, which threw us for a while.


I think that the major issues are probably resolved now, but the test suite just finished, so we'll know that a bit better once I've analysed the results in a few minutes."

[WikiWill]

There are a few positive posts from Boinc Admin in the AQUA forums suggesting new work will come out pretty soon.

"In preparation for the deployment of the next version of the AQUA app (version 4), work units are now deleted."
http://aqua.dwavesys.com/forum_thread.php?id=327&nowrap=true#4665

"I don't want to jinx it, but the unknown source of error we've been hunting for the past few weeks has finally been tracked down. It thankfully has nothing to do with our code, just one of the parameters we had never thought of changing.
...
I'm optimistic for the first time in weeks that we'll be able to get this thing out the door soon. :)"
http://aqua.dwavesys.com/forum_thread.php?id=332&nowrap=true#4659

Fingers crossed!

[Dataman]

From AQUA today

"We have been running very small test runs to make sure things are working. So far the results are promissing.

If we don't find any major problems, we will start full-scale runs today."

[veebee]

there are around 20,000 WU's available as I type this.... only short ones (2 qubit ? 2M) take about 2 mins on the i7's.

Grab 'em while they're hot !!!!