Yah, you're right. I did not pay attention to another sample rate. In the old data was 128 ms / sample, now is 500 ms.
And accordingly the old data packs were 2M samples (central-search data) or 1M samples (anti-center search) each (4 or later 10 data packs per WU).
New packaging of data is 4M samples in each, and 4 such packs in each WU.
Hmm then it turns out my computers process data much more slowly than it did in ABP2 (in terms number of samples / amounts of data per hour, for now not look to credits, although they are also slightly smaller). So the question be is - why?
In the BRP3 algorithms have been changed? Or together with new data (from other telescopes) in the same time search parameter space was increased?
And last question for today - CPU application of BRP3 have only SSE(1) version? (cuz sse2/3/4 not needed for this kind of data?) or my client get wrong version (SSE1 instead of SSE2)
Indeed, there is another parameter that is important for the runtime: the number of templates. The new data comes with a template bank of ca. 12000 templates. So for each sub-task, the radio data is matched with > 12k hypothetical binary pulsar orbits. I can't remember what the value for ABP2 was, I think around half as many templates.
There is no version for SSE2, and no version for 64 bit (which would automatically also be SSE2). I would not expect any significant performance gain by using SSE2 instead, but that's only an educated gut feeling.
There is no version for SSE2, and no version for 64 bit (which would automatically also be SSE2). I would not expect any significant performance gain by using SSE2 instead, but that's only an educated gut feeling.
CU
HBE
I wouldn't be too sure. If you remember, everyone thought the same about the regular GW app. Then, when the app was accidentally compiled as SSE2 instead of as SSE, some of us found out that SSE2 did give a pretty good performance boost over SSE.
Running Linux, I have updated my prefs, updated the project, and restarted my client. However, E@H is still sending me CPU work. Does it only work for windows?
Have updated my default preferences on the web for 'Run CPU versions of applications for which GPU versions are available' to NO. Still receiving both cuda & cpu tasks.
Message on my BOINC Client confirms its using the default preferences: 'General prefs: no separate prefs for home; using your defaults'
This is a 64bit Linux install (Ubuntu 10.10).
Do we both have the same mistake in our preferences or is there an issue with this setting for Linux?
Aem - right now, yes. The Linux CPU App is still plan-class-less. We'll fix that.
BM
Hi
a bit off topic but I am new to this site as far as Cuda is concerned.
a couple of questions
I understand ? that the new app is a 32 bit app . when will a 64 bit app be produced?
Could someone sent me a app_info file to run on a W7/Vista64 OS running 2 GTX580s.
I am familiar with app_info files as I used them with Primegrid.
A stock 580 does A WU every 37 mins or less but could do a lot better if the GPU load was higher . currently only 48/49%
My OC other 580s can do them in 34 mins but still less than 50% load.
Hope to be able to play with the settings for parallel WUs or raising the load factor.
The Primgrid Cuda app runs at 99% so it is possible for some improves with this app.
Thanks for any help
Cheers
Ross*
Aem - right now, yes. The Linux CPU App is still plan-class-less. We'll fix that.
BM
Hi
a bit off topic but I am new to this site as far as Cuda is concerned.
a couple of questions
I understand ? that the new app is a 32 bit app . when will a 64 bit app be produced?
Could someone sent me a app_info file to run on a W7/Vista64 OS running 2 GTX580s.
I am familiar with app_info files as I used them with Primegrid.
A stock 580 does A WU every 37 mins or less but could do a lot better if the GPU load was higher . currently only 48/49%
My OC other 580s can do them in 34 mins but still less than 50% load.
Hope to be able to play with the settings for parallel WUs or raising the load factor.
The Primgrid Cuda app runs at 99% so it is possible for some improves with this app.
Thanks for any help
Cheers
Ross*
RE: Yah, you're right. I
)
Indeed, there is another parameter that is important for the runtime: the number of templates. The new data comes with a template bank of ca. 12000 templates. So for each sub-task, the radio data is matched with > 12k hypothetical binary pulsar orbits. I can't remember what the value for ABP2 was, I think around half as many templates.
There is no version for SSE2, and no version for 64 bit (which would automatically also be SSE2). I would not expect any significant performance gain by using SSE2 instead, but that's only an educated gut feeling.
CU
HBE
With this new option in
)
With this new option in preferences, there still need the file app_info?
RE: There is no version for
)
I wouldn't be too sure. If you remember, everyone thought the same about the regular GW app. Then, when the app was accidentally compiled as SSE2 instead of as SSE, some of us found out that SSE2 did give a pretty good performance boost over SSE.
Now that Sandy Bridge is out,
)
Now that Sandy Bridge is out, I'd be interested to know whether compiling for AVX gives any further improvement.
Running Linux, I have updated
)
Running Linux, I have updated my prefs, updated the project, and restarted my client. However, E@H is still sending me CPU work. Does it only work for windows?
Thanks!
Having the same problem as
)
Having the same problem as telegd.
Have updated my default preferences on the web for 'Run CPU versions of applications for which GPU versions are available' to NO. Still receiving both cuda & cpu tasks.
Message on my BOINC Client confirms its using the default preferences: 'General prefs: no separate prefs for home; using your defaults'
This is a 64bit Linux install (Ubuntu 10.10).
Do we both have the same mistake in our preferences or is there an issue with this setting for Linux?
RE: Does it only work for
)
Aem - right now, yes. The Linux CPU App is still plan-class-less. We'll fix that.
BM
BM
RE: RE: Does it only work
)
Hi
a bit off topic but I am new to this site as far as Cuda is concerned.
a couple of questions
I understand ? that the new app is a 32 bit app . when will a 64 bit app be produced?
Could someone sent me a app_info file to run on a W7/Vista64 OS running 2 GTX580s.
I am familiar with app_info files as I used them with Primegrid.
A stock 580 does A WU every 37 mins or less but could do a lot better if the GPU load was higher . currently only 48/49%
My OC other 580s can do them in 34 mins but still less than 50% load.
Hope to be able to play with the settings for parallel WUs or raising the load factor.
The Primgrid Cuda app runs at 99% so it is possible for some improves with this app.
Thanks for any help
Cheers
Ross*
RE: RE: Does it only work
)
We released a new BRP3 Linux CPU App 1.05 which sports a plan class that let's you opt-out of Linux CPU work in favor of running CUDA tasks only.
Best,
Oliver
Einstein@Home Project
RE: RE: RE: Does it
)
look here http://einsteinathome.org/node/195553