Ah, so that's why. On my Windows client I have only as many WUs downloaded as are active (four at a time, one for each core). So there can't be any problems with scheduling.
... it´s increasing in steps of 0.454%, which is at me 7 to 9 minutes interval...
This is just a characteristic of FGRP2. The previous standard tasks had 20 'steps'. After each one, a point was reached where a checkpoint could be written and the displayed stats for % done and remaining time could be updated. So crunching went 5% at a time. My guess is that it's not 'convenient' to update the stats in the middle of a step.
The new tasks are 11 times longer - ie 220 steps. 100% / 220 = 0.454% so this is precisely what you are seeing. Each 'step' should be taking the same time as an old task 'step'. There's just a lot more of them :-).
The new Gamma-ray search takes between 94000 and 98000 seconds on a quad core with Vista x86, with 880 credits.
On Albert the Gamma-ray search goes via AMD GPU in about 1200 seconds. I don't know if these are the same search, but the time difference is "dramatically".
I will wait until Einstein has shorter tasks for the CPU.
The PAS on GPU is around 18000 seconds with 5000 credits. So to get RAC flowing GPU seems first choice now.
I got this task that seems to be not as long as the new FGRP2 tasks, but longer than the old ones. Is that intended? Estimated FLOPS is 51,136.
See my earlier post: "As with previous FGRP WUs, we try to cut the data sets in equally sized workunits. However there always remain a few "short ends" [for each data set]; some WUs that are significantly shorter."
We found a bug in the new App version 1.09 when processing the old, short tasks (name = "LATeah0026U...") are not affected, and the bug is fixed in App version 1.10 (out today).
Ah, so that's why. On my
)
Ah, so that's why. On my Windows client I have only as many WUs downloaded as are active (four at a time, one for each core). So there can't be any problems with scheduling.
RE: ... it´s increasing in
)
This is just a characteristic of FGRP2. The previous standard tasks had 20 'steps'. After each one, a point was reached where a checkpoint could be written and the displayed stats for % done and remaining time could be updated. So crunching went 5% at a time. My guess is that it's not 'convenient' to update the stats in the middle of a step.
The new tasks are 11 times longer - ie 220 steps. 100% / 220 = 0.454% so this is precisely what you are seeing. Each 'step' should be taking the same time as an old task 'step'. There's just a lot more of them :-).
Cheers,
Gary.
The new Gamma-ray search
)
The new Gamma-ray search takes between 94000 and 98000 seconds on a quad core with Vista x86, with 880 credits.
On Albert the Gamma-ray search goes via AMD GPU in about 1200 seconds. I don't know if these are the same search, but the time difference is "dramatically".
I will wait until Einstein has shorter tasks for the CPU.
The PAS on GPU is around 18000 seconds with 5000 credits. So to get RAC flowing GPU seems first choice now.
Greetings from
TJ
RE: I will wait until
)
You may want to focus on the BRP4 Radio Pulsar search for your CPU.
There are and will be shorter WUs for FGRP2 in the future, but I can't think of a way to specifically cherry-pick these for you.
BM
BM
RE: RE: I will wait until
)
Thanks Bernd, indeed that are the one's I was looking for but the BOINC log says that there are no tasks my PC requests. I got a couple overnight.
Greetings from
TJ
These are the run times I
)
These are the run times I have seen so far with one of my systems.
Intel 3930K @ 4.4 GHz
New tasks: 31838.55 seconds averaged
Old tasks: 2857.27 seconds averaged
Hi, I got this task that
)
Hi,
I got this task that seems to be not as long as the new FGRP2 tasks, but longer than the old ones. Is that intended? Estimated FLOPS is 51,136.
RE: I got this task that
)
See my earlier post: "As with previous FGRP WUs, we try to cut the data sets in equally sized workunits. However there always remain a few "short ends" [for each data set]; some WUs that are significantly shorter."
BM
BM
We found a bug in the new App
)
We found a bug in the new App version 1.09 when processing the old, short tasks (name = "LATeah0026U...") are not affected, and the bug is fixed in App version 1.10 (out today).
BM
BM
Any ideia on a possible
)
Any ideia on a possible release date for the GPU app? Is it weeks, months away?
I know it´s beeing tested on Albert, but can't find much information.