I am getting far more credits (1000) on the binary radio pulsar search using a GPU than on gravitational wave search on CasA using the CPU (390). Is there e reason for it? Albert@home gives me much less credits on both tasks, but there I am using only CPU, not GPU.
Tullio
I am getting far more credits (1000) on the binary radio pulsar search using a GPU than on gravitational wave search on CasA using the CPU (390). Is there e reason for it? Albert@home gives me much less credits on both tasks, but there I am using only CPU, not GPU.
Tullio
GWcasA CPU tasks and GWcasA GPU tasks receive the same 390 credits for completion/validation. BRP4G GPU tasks take accordingly longer to complete than GWcasA GPU tasks. i can't confirm it right now b/c i've got a mix of GWcasA and BRP4G tasks running on my hosts, but when i was running GWcasA GPU tasks only, or BRP4G GPU tasks only, the difference in run times between the two different types of tasks was more or less proportional to the difference in credit they earn (390 vs 1000).
In Albert@home, using only CPU, I am getting about 50 credits for a binary radio pulsar search. Here, using a GPU board, I get 1000 credits. This is what makes me think that there is something wrong in the credits system, which is often debated hotly in SETI@home.
Tullio
First of all Einstein has fixed credits for the different types of tasks and they are meant to be proportional to the amount of work in each task. As far as I know the only one here that does not follow this is FGRP work run on GPU as that app is still in development and quite inefficient right now.
Albert@home on the other hand is running another credit system, Credit New, and testing is underway to improve that system.
You are comparing different types of tasks on different projects using different credit systems.
BRP4 work for the GPU is a bundle of 16 BRP4 CPU tasks so should give you 16 times more credit than BRP4 tasks for the CPU.
BRP4 CPU task here @ Einstein gives 62.5 credit per tasks and multiplying that with 16 gives 1000 so all seems fine.
First of all Einstein has fixed credits for the different types of tasks and they are meant to be proportional to the amount of work in each task. As far as I know the only one here that does not follow this is FGRP work run on GPU as that app is still in development and quite inefficient right now.
More precise:
1. Credit for a CPU task is adjusted to be proportional to CPU-hours needed to run it. Ideally, credit/CPUh should be constant independently of the application. As the CPUs (FLOPS, caches, memory bandwidth) and applications differ quite largely, different apps behave quite differently on a given CPU, so we strive to get it right on average.
2. Credit for any task is the same, regardless of whether it is ran on a CPU or a GPU. You get the same credit for a FGRP3 or S6CasA task, regardless of whether you run it on a CPU or a GPU. A BRP4G task consists basically of 16 BRP4 tasks, and you get 16x the credit of a BRP4 (CPU) task when you run a BRP4G (GPU) task.
There is no exception to these rules. Of course the "efficiency" of the different GPU applications (and application versions) varies a lot, which is why on a give GPU you get different credit/(elapsed-)hour for different applications.
RE: Yep, we got a new bunch
)
how long is this particular batch supposed to last?
I am getting far more credits
)
I am getting far more credits (1000) on the binary radio pulsar search using a GPU than on gravitational wave search on CasA using the CPU (390). Is there e reason for it? Albert@home gives me much less credits on both tasks, but there I am using only CPU, not GPU.
Tullio
GPUs tend to be more
)
GPUs tend to be more efficient when parallel processing is possible.
RE: I am getting far more
)
GWcasA CPU tasks and GWcasA GPU tasks receive the same 390 credits for completion/validation. BRP4G GPU tasks take accordingly longer to complete than GWcasA GPU tasks. i can't confirm it right now b/c i've got a mix of GWcasA and BRP4G tasks running on my hosts, but when i was running GWcasA GPU tasks only, or BRP4G GPU tasks only, the difference in run times between the two different types of tasks was more or less proportional to the difference in credit they earn (390 vs 1000).
In Albert@home, using only
)
In Albert@home, using only CPU, I am getting about 50 credits for a binary radio pulsar search. Here, using a GPU board, I get 1000 credits. This is what makes me think that there is something wrong in the credits system, which is often debated hotly in SETI@home.
Tullio
First of all Einstein has
)
First of all Einstein has fixed credits for the different types of tasks and they are meant to be proportional to the amount of work in each task. As far as I know the only one here that does not follow this is FGRP work run on GPU as that app is still in development and quite inefficient right now.
Albert@home on the other hand is running another credit system, Credit New, and testing is underway to improve that system.
You are comparing different types of tasks on different projects using different credit systems.
BRP4 work for the GPU is a bundle of 16 BRP4 CPU tasks so should give you 16 times more credit than BRP4 tasks for the CPU.
BRP4 CPU task here @ Einstein gives 62.5 credit per tasks and multiplying that with 16 gives 1000 so all seems fine.
OK. But I got twice a 3333.00
)
OK. But I got twice a 3333.00 credit reward for a binary radio pulsar unit on GPU.
Tullio
That's BRP5 and not BRP4.
)
That's BRP5 and not BRP4. They do take quite a bit longer than BRP4 on the same GPU.
Holmis is quite correct and everything is in order with the credit awards.
Cheers,
Gary.
RE: First of all Einstein
)
More precise:
1. Credit for a CPU task is adjusted to be proportional to CPU-hours needed to run it. Ideally, credit/CPUh should be constant independently of the application. As the CPUs (FLOPS, caches, memory bandwidth) and applications differ quite largely, different apps behave quite differently on a given CPU, so we strive to get it right on average.
2. Credit for any task is the same, regardless of whether it is ran on a CPU or a GPU. You get the same credit for a FGRP3 or S6CasA task, regardless of whether you run it on a CPU or a GPU. A BRP4G task consists basically of 16 BRP4 tasks, and you get 16x the credit of a BRP4 (CPU) task when you run a BRP4G (GPU) task.
There is no exception to these rules. Of course the "efficiency" of the different GPU applications (and application versions) varies a lot, which is why on a give GPU you get different credit/(elapsed-)hour for different applications.
BM
BM
BRP4G (Arecibo data on GPUs)
)
BRP4G (Arecibo data on GPUs) ran out of work over the weekend. It will stay that way until we get new data from Arecibo.
There is a bit of Arecibo data left, but we'll use that to keep our (Android) devices fed.
Work (Parkes data) will continue to be available for the GPU applications BRP5 and later BRP6.
BM
BM