If 1 task is going slow with cuda does that mean that all the tasks with
the same start sequence ie 09ef09ad will all be vlar's and thus be slow>

---

I'd say its up to which tape it came off of. there are aa ab ac etc. the mutibeam points in multiple directions. I'd assume that if its off the same tape then it will probably have a similar angle range

---

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

I recenetly installed A Geforce 8800 GT card that BOINC version 6.4.7 recognizes as having Cuda. It sure does make A world of Difference in crunching time by using Cuda.

---

LETS BEGIN IN 2010

If 1 task is going slow with cuda does that mean that all the tasks with
the same start sequence ie 09ef09ad will all be vlar's and thus be slow>

Not certainly, but it's fairly likely. A 'tape' like 09ef09ad represents about 1.5 hours of telescope time, and observations are usually scheduled in 1 hour chunks. So when the observation is for a specific location in the sky the VLAR set may last for a full hour and that hour may be all within one 'tape'.

What is definite is that for the 107.37 seconds of a specific task, there are 256 subbands * 14 channels = 3584 WUs split. If one of those is VLAR so are all the others, and there will be at least 7168 VLAR tasks sent.

                                                               Joe

I must have got the whole tape.
I have a pile of 09fe,10fe and 12fe all are taking over
1 hr using cuda and 6.03.
Dave

---

I must have got the whole tape.
I have a pile of 09fe,10fe and 12fe all are taking over
1 hr using cuda and 6.03.
Dave

Probably observations from A2133 "The Alfa Ultra-Deep Survey: Deep HI Observations at 0<z<0.16"

Quoting a very small piece of the abstract in the descriptive PDF file:

...it is possible to achieve noise of less than 50micro-Jy with integration times of about 40 hours per pointing, ...

"Integration time" is how much total observing time they want on each target, not a continuous observation. They requested 980 hours of observing time, there's about an hour most days on the Arecibo schedules through May 7 (as far as they go now).

                                                                Joe

Thats interesting info I will read up and try to remember some of it.
6 gb on my mobo but I have trouble remembering what day it is. LOL

---

I have the GA-EX58-UD5 with a 920 on it.The sata connections go out sideways
and dont impeed the video cards.I have just built this rig about 3 weeks ago,
and it is already level with my other rig an AMD 9650 quad.
I did have 2 9600gt cards in the quad but have split them so there is 1 in each rig now

---

The VLAR effect for CUDA is related to long arrays in the pulse folding code. Just for general background, here's the maximum array size plotted against Angle Range (AR).



And here's the data set:

__AR_ MaxPoT
0.0000 32768
0.0500 32768
0.0600 27307
0.0700 23406
0.0799 20506
0.0800 40960
0.0900 36409
0.1000 32768
0.1100 29789
0.1200 27307
0.1300 25206
0.1400 23406
0.1500 21845
0.1599 20493
0.1600 40960
0.1800 36409
0.2000 32768
0.2300 28494
0.2600 25206
0.3000 21845
0.3500 18725
0.4000 16384
0.4500 14564
0.5000 13107
0.6000 10923
0.7500 8738
1.0000 6554
1.4000 4681
2.0000 3277
2.8000 2341
4.0000 1638
5.6000 1170
8.0000 819
10.000 655

Those numbers were calculated for plotting rather than absolute accuracy. If you want a really accurate maximum pulsePoTlen for a particular angle range:

1. Divide 6553.6000917504 by the angle range
2. If the result rounded to nearest integer is greater than 40960, divide the unrounded value by 2 until it is small enough.
3. The rounded integer is the array length.

If the project chose to, they could change some parameters in the WU header and make that calculation meaningless. I do not believe that's likely, most factors have been in place since the year 2000. The ALFA receiver beam width is the most recent change.

                                                                Joe

The VLAR effect for CUDA is related to long arrays in the pulse folding code. ...

Thanks for the plot, interesting.

Is the problem in that routine the amount of GPU VRAM required exceeding that physically available, or is it the way the pulse find is implimented?

ie: RAM limits or slow code?

Or is the CUDA architechture just simply unsuited for the pulse find?

Cheers,
Martin

---

See new freedom: Mageia Linux
Take a look for yourself: Linux Format
The Future is what We all make IT (GPLv3)

The VLAR effect for CUDA is related to long arrays in the pulse folding code. ...

Thanks for the plot, interesting.

Is the problem in that routine the amount of GPU VRAM required exceeding that physically available, or is it the way the pulse find is implimented?

ie: RAM limits or slow code?

Or is the CUDA architechture just simply unsuited for the pulse find?

Cheers,
Martin

It shouldn't be RAM limits, the array being analyzed consists of single floats so 40960*4 = 160 KiB max. The pattern of accessing the data is considerably different than you'd expect for graphics/video processing, though, and it's an open question whether the Nvidia guys gained a clear enough understanding of how fast folding works to produce the most efficient code.

                                                               Joe

This VLAR thing makes the tasks take so long?
When I went to 6.6.20 I reset the project and all CUDA tasks were taking ~9 minutes. 3 Days later they were taking ~2 hours. I reset the project and viola, back down to ~9 minutes. Problem is, they're creeping back into the half hour range, I have 1000 tasks in queue(don't ask) and I don't want to reset to take the hit on all the tasks, but what other solution is there?

---

This VLAR thing makes the tasks take so long?
When I went to 6.6.20 I reset the project and all CUDA tasks were taking ~9 minutes. 3 Days later they were taking ~2 hours. I reset the project and viola, back down to ~9 minutes. Problem is, they're creeping back into the half hour range, I have 1000 tasks in queue(don't ask) and I don't want to reset to take the hit on all the tasks, but what other solution is there?

Do a search for "kill VLAR"

Oh my...I don't have any tasks left!
Thanks knight.

Heres the link for everyone else.
kill vlar
There's 2 versions, one for single gpu and the other for multi gpu.
It's a .rar, so if you don't have winrar.
download.com

---