BRP4

RAMen
RAMen
Joined: 18 Jan 09
Posts: 10
Credit: 13945382
RAC: 0

I have Linux Ubuntu 64 comp

I have Linux Ubuntu 64 comp running only the GPU app: Binary Radio Pulsar Search v1.08 (BRP3cuda32nv270).
The Nvidia GPU is sitting cold. Is there an ETA for the BRP4 WU's . I like to keep it occupied but prefer to leave it here at Einstein. I need some guidance as to when I would need to switch the GPU back here. I am switching it to Collatz in the interim.

Bernd Machenschalk
Bernd Machenschalk
Moderator
Administrator
Joined: 15 Oct 04
Posts: 4267
Credit: 244931143
RAC: 16434

We are working on solving the

We are working on solving the problems with BRP4 work generation. I hope that there will be enough BRP4 tasks (and thus CUDA work) in the next few days.

BM

BM

RAMen
RAMen
Joined: 18 Jan 09
Posts: 10
Credit: 13945382
RAC: 0

RE: I hope that there will

Quote:
I hope that there will be enough BRP4 tasks (and thus CUDA work) in the next few days.

Thanks for the prompt reply will be back when you have your problems sorted

Jeroen
Jeroen
Joined: 25 Nov 05
Posts: 379
Credit: 740030628
RAC: 556

I started to get BRP4 work

I started to get BRP4 work units on my Linux system today. Thanks.

So far process time is less than what I saw with BRP3 in Linux. BRP3 took 2150-2000 seconds per work unit to process and BRP4 is taking 1950-2000 seconds per work unit to process. This is with using the nv270 app in both cases. That is a decent bump in performance compared to the BRP3 application.

Bernd Machenschalk
Bernd Machenschalk
Moderator
Administrator
Joined: 15 Oct 04
Posts: 4267
Credit: 244931143
RAC: 16434

Einstein@home is currently

Einstein@home is currently generating and sending out ~1000 BRP4 tasks per hour. I'm not sure how much more we can do. It surely is much better than before, but by far not enough to fill a buffer of tasks to send.

We're working on it.

BM

BM

Sparrow
Sparrow
Joined: 4 Jul 11
Posts: 29
Credit: 10701417
RAC: 0

RE: I started to get BRP4

Quote:

I started to get BRP4 work units on my Linux system today. Thanks.

So far process time is less than what I saw with BRP3 in Linux. BRP3 took 2150-2000 seconds per work unit to process and BRP4 is taking 1950-2000 seconds per work unit to process. This is with using the nv270 app in both cases. That is a decent bump in performance compared to the BRP3 application.

For me the speed-up was even larger. BRP3 took about 2200 seconds, BRP4 only 1700. The ratio Run time : CPU time has increased from 2:1 (BRP3) to 3:1 (BRP4).

Jeroen
Jeroen
Joined: 25 Nov 05
Posts: 379
Credit: 740030628
RAC: 556

RE: For me the speed-up was

Quote:
For me the speed-up was even larger. BRP3 took about 2200 seconds, BRP4 only 1700. The ratio Run time : CPU time has increased from 2:1 (BRP3) to 3:1 (BRP4).

Very nice. That is a large improvement in performance. What NVIDIA driver version are you running?

Sparrow
Sparrow
Joined: 4 Jul 11
Posts: 29
Credit: 10701417
RAC: 0

RE: RE: For me the

Quote:
Quote:
For me the speed-up was even larger. BRP3 took about 2200 seconds, BRP4 only 1700. The ratio Run time : CPU time has increased from 2:1 (BRP3) to 3:1 (BRP4).

Very nice. That is a large improvement in performance. What NVIDIA driver version are you running?

275.19

I just use the driver that the Ubuntu auto-update installs :-)

Betreger
Betreger
Joined: 25 Feb 05
Posts: 987
Credit: 1421815395
RAC: 779196

BRP4 is a real increase in

BRP4 is a real increase in production on my W7 machine. It uses 80% GPU vs 70%, run times reduced from 2hrs to 1hr 40min and CPU usage cut in half. If this is a new app it is good.

Bikeman (Heinz-Bernd Eggenstein)
Bikeman (Heinz-...
Moderator
Joined: 28 Aug 06
Posts: 3522
Credit: 686138689
RAC: 562245

RE: BRP4 is a real increase

Quote:
BRP4 is a real increase in production on my W7 machine. It uses 80% GPU vs 70%, run times reduced from 2hrs to 1hr 40min and CPU usage cut in half. If this is a new app it is good.

Actually it's not a new app :

md5sum einsteinbinary_BRP3_1.08_i686-pc-linux-gnu__BRP3cuda32nv270
7006bf479a4e28967f97f66e024dde69

md5sum einsteinbinary_BRP4_1.00_i686-pc-linux-gnu__BRP3cuda32nv270
7006bf479a4e28967f97f66e024dde69

The app runtime is different because the data is different (Parkes vs Arecibo), which also led to a different choice of the "template bank": the set of hypothetical orbits of a binary pulsar that are used to test for a Doppler-shifted signal in the data.

But there's hope that the CPU usage can even be cut some more by improving the code a bit.

HBE

Comment viewing options

Select your preferred way to display the comments and click "Save settings" to activate your changes.