Linux CUDA 'Special' App finally available, featuring Low CPU use

Message boards : Number crunching : Linux CUDA 'Special' App finally available, featuring Low CPU use
Message board moderation

To post messages, you must log in.

Previous · 1 . . . 64 · 65 · 66 · 67 · 68 · 69 · 70 . . . 83 · Next

AuthorMessage
Profile tullio
Volunteer tester

Send message
Joined: 9 Apr 04
Posts: 8797
Credit: 2,930,782
RAC: 1
Italy
Message 1896314 - Posted: 20 Oct 2017, 10:44:35 UTC

7.8.3 is working on my SuSE Leap 42.2 where 7.4.22 refused to work and I was obliged to run command line BOINC to SETI@home and Einstein@home. Next I shall try it on LHC@home six projects, which use VirtualBox except SixTrack.
Tullio
ID: 1896314 · Report as offensive
rob smith Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor
Volunteer moderator
Volunteer tester

Send message
Joined: 7 Mar 03
Posts: 22160
Credit: 416,307,556
RAC: 380
United Kingdom
Message 1896331 - Posted: 20 Oct 2017, 13:47:12 UTC - in response to Message 1896312.  

Updating the version of BOINC does not mean getting a new user is for either SETI or BOINC unless you do something very stupid. Nor does moving the stup from one directory to another. Your user I'd is connected to your registered email address not the working directory. Likewise if you are talking about the computer I'd then moving the data directory under Linux only needs the directory pointer modifying to the new one and the permissions confirmed, once you have stopped the old installation and installed the new version but not run it copy the old data directory over the new one and then restart.
Bob Smith
Member of Seti PIPPS (Pluto is a Planet Protest Society)
Somewhere in the (un)known Universe?
ID: 1896331 · Report as offensive
Stephen "Heretic" Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 20 Sep 12
Posts: 5557
Credit: 192,787,363
RAC: 628
Australia
Message 1896424 - Posted: 20 Oct 2017, 20:40:22 UTC - in response to Message 1896331.  
Last modified: 20 Oct 2017, 20:44:00 UTC

Updating the version of BOINC does not mean getting a new user is for either SETI or BOINC unless you do something very stupid. Nor does moving the stup from one directory to another. Your user I'd is connected to your registered email address not the working directory. Likewise if you are talking about the computer I'd then moving the data directory under Linux only needs the directory pointer modifying to the new one and the permissions confirmed, once you have stopped the old installation and installed the new version but not run it copy the old data directory over the new one and then restart.


. . I am comfortable with that on installations where everything (BOINC and the projects) are installed under the one main folder but the repo version unnerves me as it puts some things in other places and I am afraid something would be missing and trash the attempt ...

. . How sure are you that the boincdata folder is complete and self contained with the repo version?

Stephen

PS that is my forte.

??
ID: 1896424 · Report as offensive
MarkJ Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 17 Feb 08
Posts: 1139
Credit: 80,854,192
RAC: 5
Australia
Message 1896444 - Posted: 20 Oct 2017, 21:35:32 UTC - in response to Message 1896424.  
Last modified: 20 Oct 2017, 21:36:32 UTC

I am comfortable with that on installations where everything (BOINC and the projects) are installed under the one main folder but the repo version unnerves me as it puts some things in other places and I am afraid something would be missing and trash the attempt ...

. . How sure are you that the boincdata folder is complete and self contained with the repo version?

I had to upgrade all my Intel iGPU machines from Jessie to Stretch when the finally fixed the disappearing desktop issue. For each one of them I did:

1. Set BOINC to no new tasks
2. Login as root
3. Stop BOINC - in xterm type service boinc-client stop
4. Insert empty thumb drive into USB port. This brings up a window on the desktop showing the contents.
5. Using the gui copy the contents of /var/lib/boinc-client to the thumb drive. I click on the computer icon navigate to the folder, click on a single file, Ctrl-A to select all and drag it over to the thumb drive icon. Just like Windows.
5a. It complains about can’t copy sym linked files so navigate to /etc/boinc-client select all the files (there are usually four only) and drop them on the thumb drive icon as well.
6. Install new OS
7. Login as root and Install repo version of BOINC
8. Stop BOINC (in xterm type service boinc-client stop)
9. Insert thumb drive again
10. Using gui copy the files back to /var/lib/boinc-client
11. Copy the /etc files back separately using gui. They are symlinked from /var/lib. They get put there by the repo install, so overwrite them with your version from the thumb drive. I just drag and drop them one at a time.
12. In xterm cd to /var/lib/boinc-client and change permissions back to user boinc (chown boinc:boinc * -R).
13. Start it up and it resumes from where it left off.

For a new version of BOINC all I do is install it by login as root, start xterm, type apt update, type apt upgrade and it stops it, installs new version and starts it up again. You can even do it via ssh (remote login) if you want.

Hope that helps for future rebuilds.
BOINC blog
ID: 1896444 · Report as offensive
Stephen "Heretic" Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 20 Sep 12
Posts: 5557
Credit: 192,787,363
RAC: 628
Australia
Message 1896447 - Posted: 20 Oct 2017, 21:49:43 UTC - in response to Message 1896444.  

. . Thank you Mark, there is some helpful information there. I would certainly not have updated the boinc ownership otherwise.

. . I will tackle that when this trial is finished.

Stephen
ID: 1896447 · Report as offensive
Stephen "Heretic" Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 20 Sep 12
Posts: 5557
Credit: 192,787,363
RAC: 628
Australia
Message 1896450 - Posted: 20 Oct 2017, 22:02:18 UTC

. . Hi TBar

. . Well I have the first inconclusive results from CUDA60 3v. One is a normal end task and the other is a very late overflow. The late overflow already has 3 results and is waiting on a 4th.

. . I will keep an eye on them.

. . The results are the ones dated 20th Oct 21:03:56

http://setiathome.berkeley.edu/results.php?hostid=8222433&offset=0&show_names=0&state=3&appid=

Stephen

??
ID: 1896450 · Report as offensive
Profile Jeff Buck Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor
Volunteer tester

Send message
Joined: 11 Feb 00
Posts: 1441
Credit: 148,764,870
RAC: 0
United States
Message 1896517 - Posted: 21 Oct 2017, 3:49:40 UTC

Here's another Inconclusive involving the zi3xs3 app, very similar to the one I posted 3 days ago (Message 1895798). There's a difference in the order that the Spikes and Autocorrs are interleaved in the Stderr, though all can pretty much be matched up until the very last reported signal when the overflow condition was triggered. Due to that different reporting order, the v8.08 (alt) app reported a Spike as the last signal, whereas the zi3xs3 app reported an Autocorr, thus leading to the Inconclusive result.

I don't recall seeing this particular Spike vs. Autocorr issue occurring prior to the zi3xs3 version, although perhaps I just overlooked it. It seems like just a minor change in reporting sequence would probably bring the Special App in line with the others.

Workunit 2716057052 (01ap07aa.20141.11524.9.36.255)
Task 6103556320 (S=23, A=7, P=0, T=0, G=0, BG=2.864089) x41p_zi3xs3, Cuda 9.00 special
Task 6104162821 (S=24, A=6, P=0, T=0, G=0, BG=2.864112) v8.08 (alt) windows_x86_64

One of my hosts should be running the tiebreaker overnight with SoG, which I expect will match up with v8.08 (alt). Since the results are so close, I expect Petri's host will get credit, too, as it did with the similar WU from 3 days ago.
ID: 1896517 · Report as offensive
Stephen "Heretic" Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 20 Sep 12
Posts: 5557
Credit: 192,787,363
RAC: 628
Australia
Message 1896572 - Posted: 21 Oct 2017, 8:12:04 UTC

. . Before pulling the GT730 out I decided to run it with blocking sync off. I can safely say ... nada ... zip not a bit of difference. The only change on this card by turning it off was not being able to crunch on one CPU core. Otherwise the run times are virtually identical.

. . So if anyone wants to run CUDA60 on a GT730 let them know not to bother turning it off. The rig will lose 10 to 20 WU per day from the CPU and be lucky to gain one on the GPU. Otherwise the run was reasonably successful. Nett result was about a 10 to 15 % improvement over CUDA50 and Windows (running singles, about half that over CUDA50 running doubles)

Stephen

..
ID: 1896572 · Report as offensive
TBar
Volunteer tester

Send message
Joined: 22 May 99
Posts: 5204
Credit: 840,779,836
RAC: 2,768
United States
Message 1896609 - Posted: 21 Oct 2017, 13:11:00 UTC - in response to Message 1896572.  
Last modified: 21 Oct 2017, 13:47:06 UTC

...Nett result was about a 10 to 15 % improvement over CUDA50 and Windows (running singles, about half that over CUDA50 running doubles)
Are you sure about that 10 to 15%? I haven't looked at the 730 much, but the one I found at beta is showing much more than 10 - 15%, https://setiweb.ssl.berkeley.edu/beta/result.php?resultid=28861148
Windows AR 2.7 = 47 min 13 sec
Special AR 1.5 = 7 min 50 sec
https://setiathome.berkeley.edu/result.php?resultid=6106162921
I Dunno, is the one at beta just really slow...or running Multiples?
Maybe I need more coffee...

Found another one at Beta, it hasn't run any CUDAs lately, but the Details show it running OpenCL faster than CUDA 50 and the Special App is much faster than his OpenCL times, https://setiweb.ssl.berkeley.edu/beta/results.php?hostid=72029&offset=240
ID: 1896609 · Report as offensive
Stephen "Heretic" Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 20 Sep 12
Posts: 5557
Credit: 192,787,363
RAC: 628
Australia
Message 1896644 - Posted: 21 Oct 2017, 16:58:40 UTC - in response to Message 1896609.  

...Nett result was about a 10 to 15 % improvement over CUDA50 and Windows (running singles, about half that over CUDA50 running doubles)
Are you sure about that 10 to 15%? I haven't looked at the 730 much, but the one I found at beta is showing much more than 10 - 15%, https://setiweb.ssl.berkeley.edu/beta/result.php?resultid=28861148
Windows AR 2.7 = 47 min 13 sec
Special AR 1.5 = 7 min 50 sec
https://setiathome.berkeley.edu/result.php?resultid=6106162921
I Dunno, is the one at beta just really slow...or running Multiples?
Maybe I need more coffee...


. . Firstly there are several breeds of the GT730. The older chipset (192 CUDA cores) with the 128 bit memory bus which is a slower GPU with 1GB DDR3 ram, then the improved chipset (384 CUDA cores) with a 64 bit bus (I know one step forward and one step back) and 1GB of DDR3 ram. Then there is mine, the improved chipset with the 64bit bus and 2 GB DDR5 ram. It is a quicker card overall. Memory bandwidth goes as follows 28Mbit, 17Mbits and then mine at 40Mbits.

. . Secondly, the task you pulled numbers on is a halfling (AR=2.5 something) so to take 47 mins with CUDA50 he would have to be running doubles on a slower GT730 (his has the standard clock of 902MHz, mine is factory overclocked to 1006MHz). I will give you my numbers at the end.

. . Thirdly the Special sauce task is also a halfling (AR 1.5 something) and running singly but seems to be on a different card as it has 2GB ram not 1GB as in the first instance. The few halflings I have run made similar times with CUDA60 ver 3v.

Found another one at Beta, it hasn't run any CUDAs lately, but the Details show it running OpenCL faster than CUDA 50 and the Special App is much faster than his OpenCL times, https://setiweb.ssl.berkeley.edu/beta/results.php?hostid=72029&offset=240


. . This is also a 2GB board with standard clocking. But running different apps and most of the non GBT tasks seem to be halflings as well. The funny thing is that GBT tasks on these cards run quicker with SoG than with CUDA50, the down side if Arecibo tasks are MUCH slower with SoG than with CUDA50. My comparisons are based on the task type I have had the most of during this short test sample, that is NARAs (normal AR Arecibo)

My numbers:

CUDA60 Ver 3v

BS on

blc04 x 1 @ 37.8
NARA x 13 @ 25.5 to 27.0
VHAR x 6 @ 7.2 to 7.8

BS off

blc02 x 1 @ 57.2
blc04 x 6 @37.2 to 37.7
NARA x 17 @ 25.6 to 26.4
VHAR x 0

CUDA50 (Windows)

blc no stats to speak of as I rescheduled them to the CPU as much as I could
NARA x {heaps} @ 54.0 to 61.0 (doubles so 27 to 31 per task)
NARA x {not quite as many} @ 32 to 35 {singles}

. . So no discernible improvement by disabling BS. And only a few mins quicker than CUDA50. At best 4 - 5 mins, at worst 2 - 3, or if running only singles then maybe as good as 5 or 8. Definitely an improvement but not in the league of more modern cards.

. . I hope that answers your questions without boring you too much.

Stephen

:)
ID: 1896644 · Report as offensive
dallasdawg

Send message
Joined: 19 Aug 99
Posts: 49
Credit: 142,692,438
RAC: 2
United States
Message 1896652 - Posted: 21 Oct 2017, 17:36:15 UTC - in response to Message 1895647.  
Last modified: 21 Oct 2017, 18:01:52 UTC

ignore - answers already in thread
ID: 1896652 · Report as offensive
TBar
Volunteer tester

Send message
Joined: 22 May 99
Posts: 5204
Credit: 840,779,836
RAC: 2,768
United States
Message 1896667 - Posted: 21 Oct 2017, 20:06:58 UTC - in response to Message 1896644.  
Last modified: 21 Oct 2017, 20:12:43 UTC

CUDA50 (Windows)

NARA x {heaps} @ 54.0 to 61.0 (doubles so 27 to 31 per task)
NARA x {not quite as many} @ 32 to 35 {singles}

. . So no discernible improvement by disabling BS. And only a few mins quicker than CUDA50. At best 4 - 5 mins, at worst 2 - 3, or if running only singles then maybe as good as 5 or 8. Definitely an improvement but not in the league of more modern cards.

. . I hope that answers your questions without boring you too much....


Sorry, I'm just having a Really hard time with those numbers. In Every case I've seen the Special App is almost Twice as fast as CUDA 50 or OpenCL. There are quite a few 730s out there just like yours that have much different numbers than you are claiming, here's another one, https://setiathome.berkeley.edu/results.php?hostid=8241817&offset=140 From the report times it's clear he's running one at a time. I also remember when the SoG App was first released it ran very badly on the 730s and many people with 730s were switching over to CUDA 50. So, we have a good base to relate to, later the SoG App was improved to where it's better now. I suppose it could be something with your particular 730, or perhaps you're not reading your notes correctly again? In any event, I can't buy CUDA 50 is almost as fast as the Special App.
ID: 1896667 · Report as offensive
Stephen "Heretic" Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 20 Sep 12
Posts: 5557
Credit: 192,787,363
RAC: 628
Australia
Message 1896763 - Posted: 22 Oct 2017, 2:20:30 UTC - in response to Message 1896667.  
Last modified: 22 Oct 2017, 2:56:04 UTC


Sorry, I'm just having a Really hard time with those numbers. In Every case I've seen the Special App is almost Twice as fast as CUDA 50 or OpenCL. There are quite a few 730s out there just like yours that have much different numbers than you are claiming, here's another one, https://setiathome.berkeley.edu/results.php?hostid=8241817&offset=140 From the report times it's clear he's running one at a time. I also remember when the SoG App was first released it ran very badly on the 730s and many people with 730s were switching over to CUDA 50. So, we have a good base to relate to, later the SoG App was improved to where it's better now. I suppose it could be something with your particular 730, or perhaps you're not reading your notes correctly again? In any event, I can't buy CUDA 50 is almost as fast as the Special App.


. . My apologies in going off at a bit of a tangent below but that rig is definitely running doubles. Either that or his GPU is on prozac. And I estimate the RAC on this machine if I had continued with CUDA60 would have settled somewhere between 7K and 8K at best, probably more realistically about mid to high 6K. When I was running the 730 my RAC was 5.3K.


. . OK, as Bruce and I have just been doing testing for you for Kepler cards I didn't think there would be many results for CUDA60 special on other cards. But I am prepared to believe/accept that they might be similar to those I have observed with my card. One thing I have noticed is that when crunching with the more modern apps, especially the special sauce variety, there is little real world benefit from the higher level hardware for the individual series. The run times on my 1050ti are almost indistinguishable from the run times on my 1050 despite having an extra CU (6 vs 5) and 128 more cuda cores, and even with SoG the run times Wiggo has been getting on his 1060-3GB are almost the same as those I was getting from my 1060-6GBs with 10 CUs over his 9 and lots more CUDA cores in the mix. But I get the impression you doubt the performance I was getting from the GT 730 before I retired it. If you want to fly to OZ you can read my logs :)

. . They were as stated and very very consistent. If you want to understand why the numbers you keep dragging out seem to contradict that please check them and eliminate GT730's running VLAR tasks under CUDA50, they will obviously be bad. and useless when comparing to the NARAs under CUDA60. And yes the halflings do run very fast with the special app. Very fast indeed taking 7 to 8 mins compared to 25 to 27 for the NARAs. But certainly, for any GT730 users out there with slower cards and little or no optimisation on their rigs they would see a bigger improvement in run times by changing to CUDA60 than I have. But I can only cite the figures I have from my real world results.

. . Another thing about the seemingly poor CUDA50 performance on the 730s running multiple tasks with a mix of GBT tasks is that when GPUs are shared by a GBT task and an Arecibo task, the run times for the GBT task will be somewhat lower, but the run times for the Arecibo task will be wwaaaayyyy longer. My numbers are for optimised processing with task segregation. Yes I'm a taskist. :) But the few SoG Arecibo tasks I ran on the 730 took typically 47 to about 70 mins (from Bloc02 to Bloc05 tasks) {as singles, doubles were far, far slower} and the sole Bloc02 I ran with CUDA60 took 57 mins, not an improvement at all in my opinion. But if you compare my times with Bloc04 CUDA60 against the times I had using SoG (an earlier version than 8.20/8.22) it is about 37 mins compared to about 50 mins, but the current batch of Blc04 tasks are a wonderful exception and on my other i5 rig are achieving identical run times with Arecibo WUs.

. . The big advantage with the special app is it is best at running singles and that effect is eliminated.

. . So I guess the bottom line is, you are right, for a large percentage of the relatively few GT730 crunchers out there they would find a bigger improvement in their productivity by changing to CUDA60 than I have achieved. But for me since this card is retired and going to another home the point is moot. I am very happy with the 1050ti running CUDA80 special. Now that really does see a halving of run times (better than half actually).

Stephen

:)
ID: 1896763 · Report as offensive
Profile Jeff Buck Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor
Volunteer tester

Send message
Joined: 11 Feb 00
Posts: 1441
Credit: 148,764,870
RAC: 0
United States
Message 1896776 - Posted: 22 Oct 2017, 3:23:03 UTC
Last modified: 22 Oct 2017, 3:24:05 UTC

Here's an Inconclusive where a Cuda 9.00 app disagrees with one of the earlier Cuda 8.00 Special Apps. Both reported 30 Spikes but the Spikes are totally different. Where all the Spikes the x41p_zi3t2b on my host reported have an fft_len of 2k or 4k, all the ones from the x41p_zi3x app have an fft_len of 256 or 512.

Workunit 2717793652 (02ap07aa.5161.14796.14.41.0)
Task 6107177460 (S=30, A=0, P=0, T=0, G=0, BG=0) x41p_zi3x, Cuda 9.00 special
Task 6107177461 (S=30, A=0, P=0, T=0, G=0, BG=0) x41p_zi3t2b, Cuda 8.00 special

I'm actually hoping that the tiebreaker, which looks like it will run with the stock Windows CPU app, agrees with Laurent's Cuda 9.00 host and not my host. That should give some good evidence that the newer app is better at matching up with the "gold standard" (at least on this type of -9 overflow).
ID: 1896776 · Report as offensive
TBar
Volunteer tester

Send message
Joined: 22 May 99
Posts: 5204
Credit: 840,779,836
RAC: 2,768
United States
Message 1896789 - Posted: 22 Oct 2017, 5:39:31 UTC - in response to Message 1896763.  
Last modified: 22 Oct 2017, 5:46:06 UTC

Stephen, this is the problem with your post,
CUDA50 (Windows)
blc no stats to speak of as I rescheduled them to the CPU as much as I could
NARA x {heaps} @ 54.0 to 61.0 (doubles so 27 to 31 per task)
NARA x {not quite as many} @ 32 to 35 {singles}

Now, from my experience, it takes a 730 about an Hour to run a AR 0.40 Arecibo task, I've already shown you 3 examples...there are more.
The Special App runs an AR 0.40 in about half that time;
Workunit 2717313444
Run time 25 min 39 sec
WU true angle range is : 0.403722
This agrees with all the other results in the Known Universe. The Special App is around twice as fast.

Can you show me One good example of a 730 running an AR 0.40 Arecibo task in cuda 50 twice as fast as everyone else?
The results would be perfectly normal at "NARA x {heaps} @ 54.0 to 61.0" I'm just not buying your GPU is twice as fast running cuda 50 , no matter how much you write.
BTW, we go by true angle range around here, been that way a long time.
ID: 1896789 · Report as offensive
Stephen "Heretic" Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 20 Sep 12
Posts: 5557
Credit: 192,787,363
RAC: 628
Australia
Message 1896804 - Posted: 22 Oct 2017, 8:29:05 UTC - in response to Message 1896789.  
Last modified: 22 Oct 2017, 8:39:44 UTC

Stephen, this is the problem with your post,
CUDA50 (Windows)
blc no stats to speak of as I rescheduled them to the CPU as much as I could
NARA x {heaps} @ 54.0 to 61.0 (doubles so 27 to 31 per task)
NARA x {not quite as many} @ 32 to 35 {singles}



. . It wasn't a problem for me I was very happy with the results. Maybe Wiggo or Grant can recall if they took any notice of what that machine was doing at the time and corroborate the facts. Otherwise you would need to be able to access the records for the WU's I processed 12 months ago. And the halflings would take about 18 to 20 mins run singly.

Now, from my experience, it takes a 730 about an Hour to run a AR 0.40 Arecibo task, I've already shown you 3 examples...there are more.


. . The examples you cited were either halflings with ARs of 1.5 or 2.5 or VLARs. But have you asked the owners of the rigs involved if they are running singles or doubles? You say they have run times of about an hour or more, and that is the run times I was getting running doubles. But hey, maybe my little GT730 is hopped up on ice or something :) Either way, I repeat, it is a moot point. This unit will remain out of SETI production as it is moving house.

The Special App runs an AR 0.40 in about half that time;
Workunit 2717313444
Run time 25 min 39 sec
WU true angle range is : 0.403722


. . Well we can agree on the run times using the CUDA60 on normal AR arecibo tasks.

This agrees with all the other results in the Known Universe.


. . That statement is completely 100% false ...

The Special App is around twice as fast.
Can you show me One good example of a 730 running an AR 0.40 Arecibo task in cuda 50 twice as fast as everyone else?
The results would be perfectly normal at "NARA x {heaps} @ 54.0 to 61.0" I'm just not buying your GPU is twice as fast running cuda 50 , no matter how much you write.
BTW, we go by true angle range around here, been that way a long time.


. . Only if you can access the records for WUs processed prior to last Christmas, because that was when the 730 was retired and superseded by the 1050ti. But even those WUs were doubles. If you want singles run times you would need to go back months earlier. Does your mention of "true angle range" mean you don't believe the AR recorded in the stderr files? That is the only measure I have of AR so I can't imagine what you are alluding to with that comment. It's a shame this topic wasn't happening 12 months ago, then I could have referred you to task numbers. :)

Stephen

:)
ID: 1896804 · Report as offensive
TBar
Volunteer tester

Send message
Joined: 22 May 99
Posts: 5204
Credit: 840,779,836
RAC: 2,768
United States
Message 1896805 - Posted: 22 Oct 2017, 8:36:56 UTC - in response to Message 1896804.  
Last modified: 22 Oct 2017, 8:42:15 UTC

I see, you can't backup your claims. So, I'm going to have to call B.S.
Please don't post B.S. claims in my thread. I don't appreciate it.
ID: 1896805 · Report as offensive
Stephen "Heretic" Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 20 Sep 12
Posts: 5557
Credit: 192,787,363
RAC: 628
Australia
Message 1896806 - Posted: 22 Oct 2017, 8:50:33 UTC - in response to Message 1896805.  

I see, you can't backup your claims. So, I'm going to have to call B.S.
Please don't post B.S. claims in my thread. I don't appreciate it.


. . Host ID 7972413 WUs processed on 27th to 29th August 2016

. . Access those records and there is your allegedly "none-existent" proof. So far the only BS is from your side.

Stephen
ID: 1896806 · Report as offensive
TBar
Volunteer tester

Send message
Joined: 22 May 99
Posts: 5204
Credit: 840,779,836
RAC: 2,768
United States
Message 1896809 - Posted: 22 Oct 2017, 8:58:22 UTC - in response to Message 1896806.  
Last modified: 22 Oct 2017, 9:17:29 UTC

Show me an example of your claims cuda 50 is almost as fast as the Special App, until then it's B.S.
Here's another one showing about an hour for an AR 0.40, https://setiathome.berkeley.edu/results.php?hostid=8055793&offset=80
There are a few more. I can produce many, I'm still waiting for you to produce 1.
Here's another, https://setiathome.berkeley.edu/results.php?hostid=6691604&offset=320
https://setiathome.berkeley.edu/results.php?hostid=8154163&offset=40
https://setiathome.berkeley.edu/results.php?hostid=8086970&offset=80
ID: 1896809 · Report as offensive
Profile Jeff Buck Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor
Volunteer tester

Send message
Joined: 11 Feb 00
Posts: 1441
Credit: 148,764,870
RAC: 0
United States
Message 1896912 - Posted: 22 Oct 2017, 18:41:37 UTC - in response to Message 1896776.  

Here's an Inconclusive where a Cuda 9.00 app disagrees with one of the earlier Cuda 8.00 Special Apps. Both reported 30 Spikes but the Spikes are totally different. Where all the Spikes the x41p_zi3t2b on my host reported have an fft_len of 2k or 4k, all the ones from the x41p_zi3x app have an fft_len of 256 or 512.

Workunit 2717793652 (02ap07aa.5161.14796.14.41.0)
Task 6107177460 (S=30, A=0, P=0, T=0, G=0, BG=0) x41p_zi3x, Cuda 9.00 special
Task 6107177461 (S=30, A=0, P=0, T=0, G=0, BG=0) x41p_zi3t2b, Cuda 8.00 special

I'm actually hoping that the tiebreaker, which looks like it will run with the stock Windows CPU app, agrees with Laurent's Cuda 9.00 host and not my host. That should give some good evidence that the newer app is better at matching up with the "gold standard" (at least on this type of -9 overflow).
The tiebreaker hasn't reported yet, but I just ran a bench with setiathome_8.00_windows_intelx86 here on my daily driver. Unfortunately, the 30 spikes that show up in the result file match with the zi3t2b, Cuda 8.00 result, not the x41p_zi3x, Cuda 9.00 result, so it appears that the zi3x has gone the wrong way. It might be useful if anyone else can test this WU offline with some other apps to see if there's a consensus.
ID: 1896912 · Report as offensive
Previous · 1 . . . 64 · 65 · 66 · 67 · 68 · 69 · 70 . . . 83 · Next

Message boards : Number crunching : Linux CUDA 'Special' App finally available, featuring Low CPU use


 
©2024 University of California
 
SETI@home and Astropulse are funded by grants from the National Science Foundation, NASA, and donations from SETI@home volunteers. AstroPulse is funded in part by the NSF through grant AST-0307956.