The new BRP6 application version 1.47 has been published for Beta test. It features the optimizations from HBE as announced e.g. here, and the "latest and greatest" BOINC API code, including the fix mentioned by Richard e.g. here.
BM
I spotted beta work this morning, and artificially downloaded some extra, and suspended non-beta in queue to get an early look. On two of three hosts all available beta tasks errored out in sequence, after about 2.5 seconds of run time, with errors first asserting stack overflow with too deep recursion, then claiming no suitable CUDA device. The third host is currently apparently successfully running a full set (two each for two dissimilar GPUs) and at 10 to 15 percent completion is reporting progress consistent with an appreciable speedup. (current indications suggest completion time at roughly 70% of previous application on a host running a GTX 750 and a GTX 660, both at 2X--but too soon to tell).
The run time of the beta app should be a bit more data-dependent than with the current official app, so there will be quite a spread in speed-up for different WUs, but 70% of previous runtime is in the ballpark we are expecting. A later version with CUDA 5.5 should cut another (very roughly) 10% off the run-time.
I spotted beta work this morning, and artificially downloaded some extra, and suspended non-beta in queue to get an early look. On two of three hosts all available beta tasks errored out in sequence, after about 2.5 seconds of run time ...
Same happend to me with one host: GTX670 / 347.52
On the two other hosts: GTX460 and GTX560Ti / 347.52 BRP6-Beta is working normal, I think ...
I know I am a part of a story that starts long before I can remember and continues long beyond when anyone will remember me [Danny Hillis, Long Now]
I just configured a host with a HD7850 to get BRP6 1.47 beta tasks. I had pre-deployed the 1.47 Beta-opencl-ati app which was duly found and the download of that file was skipped (as expected). However the file einsteinbinary_BRP6_1.40_x86_64-pc-linux-gnu__BRP6-Beta-opencl-ati was also downloaded at the same time as the 1.47 version was being skipped.
Why are there two separate beta versions of the app being required? The downloaded tasks show as 1.47 in the app column of BOINC Manager.
This is definitely a bug, not a feature. I really don't know how the 1.40 binary slipped into the 1.47 app bundle, but this was unintentional. It is certainly not require for the app's operation.
I know that the runtime should increase compared to BRP5, but 20hrs for BRP6 compared to 4.5h for BRP5 seems a bit too much. Does anyone have a similar problem? or it is problem with my machine?
It has E8600 CPU and Radeon 5770 GPU . Other my hosts have reasonable runtime difference between BRP5 and BRP6
A few subjective and some quantitative observations on my one host which is successfully running the beta tasks:
Running at 2X on both a 750 Ti and a 660, the temperature reported on the 660 is several degrees higher than with the non-beta application running Parkes. Also it appears that the GPU-Z reported GPU load percent is up appreciably. The CPU time consumed by the support application is down dramatically (both as a fraction of CPU in current use during the run, and yet more strongly for total CPU time consumed for a WU. I'm sure power consumption is up as well, but don't have proper before/after comparisons on any of these four observations.
All of these are strongly favorable indications that the new application is making more effective use of these two GPUs on this work.
However the temperature/power increase could push installations running near their limits over some edge. In my personal case I run a TThrottle configuration intended to reduce power consumption in hotter weather, and to shift relative power consumption toward the cooler parts of the day. It is winter here, and at my settings in force my GTX 660 was not breaching the temperature limit to initiate any throttling. I spotted it throttling by 3% when the run was perhaps a quarter of the way completed and raised the limits, but this will have slightly pessimized the first completion timings. My very first completed result completed on the GTX 660 running 2X (with another beta as companion) in a reported 3:15:09 elapsed time, with a remarkably low 0:15:23 reported CPU time. For comparison, the non-Beta Parkes GTX 660 results finishing in the previous day ran about 4:41:00 elapsed time and 1:01:00 CPU time. However, of my pair of simultaneous jobs, this one ran a bit faster, and substantially lower CPU time than the companion. This may reflect my failure yet to update Process Lasso settings to handle the CPU support task for the beta Parkes work the same as I have it set for the standard work.
While I was typing, the companion ask finished, so in summary for two tasks I have so far seen:[pre]parameter beta1 beta2 non-beta average
Elapsed 3:15:09 3:22:33 4:41:00
CPU time 0:15:23 0:25:21 1:00:40 [/pre]
Obviously the GTX 660 is non-Maxwell. I have hopes my Maxwells may speed up by a somewhat higher fraction, somewhat born out by mid-course readouts on two 750 ti tasks.
I don't seem to get the new 1.47 application. Is there one for linux yet?
However I'm only getting BRP4G since midday.
Beta option is activated and I'm getting beta workunits for the CPU.
My client can't download 1.47 opencl-ati application for Linux. I keep getting:
2/27/2015 7:17:59 PM | Einstein@Home | Started download of einsteinbinary_BRP6_1.40_x86_64-pc-linux-gnu__BRP6-Beta-opencl-ati
2/27/2015 7:18:01 PM | Einstein@Home | Giving up on download of einsteinbinary_BRP6_1.40_x86_64-pc-linux-gnu__BRP6-Beta-opencl-ati: permanent HTTP error
(it's a 404 - not found error)
and all 1.47 tasks fail with download error.
RE: 2015-02-27
)
I also improved the scheduler logging a bit to also show the version number and plan class for the selected app version now, like:
(the app version id is pretty useless to anyone out there)
BM
BM
RE: The new BRP6
)
I spotted beta work this morning, and artificially downloaded some extra, and suspended non-beta in queue to get an early look. On two of three hosts all available beta tasks errored out in sequence, after about 2.5 seconds of run time, with errors first asserting stack overflow with too deep recursion, then claiming no suitable CUDA device. The third host is currently apparently successfully running a full set (two each for two dissimilar GPUs) and at 10 to 15 percent completion is reporting progress consistent with an appreciable speedup. (current indications suggest completion time at roughly 70% of previous application on a host running a GTX 750 and a GTX 660, both at 2X--but too soon to tell).
I've started a "trouble" thread in Problems and Bug Reports.
Thanks. The run time of
)
Thanks.
The run time of the beta app should be a bit more data-dependent than with the current official app, so there will be quite a spread in speed-up for different WUs, but 70% of previous runtime is in the ballpark we are expecting. A later version with CUDA 5.5 should cut another (very roughly) 10% off the run-time.
HBE
RE: RE: I spotted beta
)
Same happend to me with one host: GTX670 / 347.52
On the two other hosts: GTX460 and GTX560Ti / 347.52 BRP6-Beta is working normal, I think ...
I know I am a part of a story that starts long before I can remember and continues long beyond when anyone will remember me [Danny Hillis, Long Now]
I just configured a host with
)
I just configured a host with a HD7850 to get BRP6 1.47 beta tasks. I had pre-deployed the 1.47 Beta-opencl-ati app which was duly found and the download of that file was skipped (as expected). However the file einsteinbinary_BRP6_1.40_x86_64-pc-linux-gnu__BRP6-Beta-opencl-ati was also downloaded at the same time as the 1.47 version was being skipped.
Why are there two separate beta versions of the app being required? The downloaded tasks show as 1.47 in the app column of BOINC Manager.
Cheers,
Gary.
This is definitely a bug, not
)
This is definitely a bug, not a feature. I really don't know how the 1.40 binary slipped into the 1.47 app bundle, but this was unintentional. It is certainly not require for the app's operation.
BM
BM
Hi all, I have a wierd
)
Hi all,
I have a wierd performance problem with BRP6 v1.41 on one of my machines
http://einsteinathome.org/host/11738443/tasks
I know that the runtime should increase compared to BRP5, but 20hrs for BRP6 compared to 4.5h for BRP5 seems a bit too much. Does anyone have a similar problem? or it is problem with my machine?
It has E8600 CPU and Radeon 5770 GPU . Other my hosts have reasonable runtime difference between BRP5 and BRP6
A few subjective and some
)
A few subjective and some quantitative observations on my one host which is successfully running the beta tasks:
Running at 2X on both a 750 Ti and a 660, the temperature reported on the 660 is several degrees higher than with the non-beta application running Parkes. Also it appears that the GPU-Z reported GPU load percent is up appreciably. The CPU time consumed by the support application is down dramatically (both as a fraction of CPU in current use during the run, and yet more strongly for total CPU time consumed for a WU. I'm sure power consumption is up as well, but don't have proper before/after comparisons on any of these four observations.
All of these are strongly favorable indications that the new application is making more effective use of these two GPUs on this work.
However the temperature/power increase could push installations running near their limits over some edge. In my personal case I run a TThrottle configuration intended to reduce power consumption in hotter weather, and to shift relative power consumption toward the cooler parts of the day. It is winter here, and at my settings in force my GTX 660 was not breaching the temperature limit to initiate any throttling. I spotted it throttling by 3% when the run was perhaps a quarter of the way completed and raised the limits, but this will have slightly pessimized the first completion timings. My very first completed result completed on the GTX 660 running 2X (with another beta as companion) in a reported 3:15:09 elapsed time, with a remarkably low 0:15:23 reported CPU time. For comparison, the non-Beta Parkes GTX 660 results finishing in the previous day ran about 4:41:00 elapsed time and 1:01:00 CPU time. However, of my pair of simultaneous jobs, this one ran a bit faster, and substantially lower CPU time than the companion. This may reflect my failure yet to update Process Lasso settings to handle the CPU support task for the beta Parkes work the same as I have it set for the standard work.
While I was typing, the companion ask finished, so in summary for two tasks I have so far seen:[pre]parameter beta1 beta2 non-beta average
Elapsed 3:15:09 3:22:33 4:41:00
CPU time 0:15:23 0:25:21 1:00:40 [/pre]
Obviously the GTX 660 is non-Maxwell. I have hopes my Maxwells may speed up by a somewhat higher fraction, somewhat born out by mid-course readouts on two 750 ti tasks.
I don't seem to get the new
)
I don't seem to get the new 1.47 application. Is there one for linux yet?
However I'm only getting BRP4G since midday.
Beta option is activated and I'm getting beta workunits for the CPU.
My client can't download 1.47
)
My client can't download 1.47 opencl-ati application for Linux. I keep getting:
(it's a 404 - not found error)
and all 1.47 tasks fail with download error.