Until the first Einstein WU validates, we can't see. Over on Albert, they awarded 2000 for a couple I saw, but I doubt that has much predictive value. Possibly the project may use data from work returned before the validator is enabled to help decide on initial credit award. From previous behavior, they may adjust that level if more data suggest it is far from alignment.
The runtime difference between both systems is larger with BRP5 than BRP4. I have to look and see why the second system is running the BRP5 tasks significantly slower than my first system.
So far all completed tasks are pending validation.
I had just last night switched from Catalysts 12.1 to 13.4 on my HD6850, and I was checking if they'd run any faster, or slower than under the previous drivers.
Not much between 12.1 and 13.4 in times, on average still around 2,300 seconds. But then in came a couple of BRP5s, which run at 17,600 seconds. Ouch.
For the moment I disabled BRP5 as an option, they'll just have to wait until next month, when I'm planning to get me a HD7850 or HD7870. Probably the latter.
I'm wondering why BRP5 tasks have been put into queue, but their validator isn't running ?
I'm sure there is a reason, but I'd really like to see it turned-on. I'd like to know if the work is being done successfully.
I've got about 100 of them waiting to validate, some with NVIDA some with AMD, several different OSes. If I could get one per machine to validate I'd be happy.
I'm wondering why BRP5 tasks have been put into queue, but their validator isn't running ?
- BRP5 tasks were issued because we are running short of BRP4 data (the server status page shows when the remaining data will be finished i.e. tasks returned, it doesn't show when we will run out of tasks to send out).
- Starting BRP5 was actually delayed, mainly because of me being sick most of last week.
- We started shipping BRP5 on Friday. As no one would be able to watch what goes right or wrong over the weekend, we decided to deal with validation (assimilation etc.) on Monday, when we could easily watch and possibly correct errors. Also too we'll have a better understanding of the run-time relation between BRP4 and BRP5 than what we got from the very short tests on Albert.
We're somewhat victims of our own success. The recent increase in GPU computing power led to a higher processing rate of Arecibo data than the rate at which Arecibo can capture it, and also to tasks run-times so short that the throughput began to seriously stress our DB. BRP5 is targeted to address both issues.
According to my short observations and with assumption that the results are correct, the individual task should give about 5000 to 8000 points. I also think it would be good to have a separate option to set the GPU utilization factor for BRP4 and BRP5.
Gattorantolo wrote:How many
)
Until the first Einstein WU validates, we can't see. Over on Albert, they awarded 2000 for a couple I saw, but I doubt that has much predictive value. Possibly the project may use data from work returned before the validator is enabled to help decide on initial credit award. From previous behavior, they may adjust that level if more data suggest it is far from alignment.
I've had a look at the
)
I've had a look at the results from the first day of processing and its interesting. The following table is a summary:
BRP5 workload compared to BRP4 tasks has increased by a factor of ~5x, but the processing times are ~10x and CPU loading is ~15x.
I am already seeing a huge impact on the number of total wu processed per day across all hosts.
Gord
Here are my results running
)
Here are my results running three tasks per GPU.
Per task total runtime, overall GPU load, and power draw if available
System #1 - 2x 7970
BRP4: 1360 seconds / 95% GPU Load / 600w
BRP5: 12,100 seconds / 92% GPU Load / ?
System #2 - 3x 7970
BRP4: 1450 seconds / 95 % GPU Load / 780w
BRP5: 14,100 seconds / 90 % GPU Load / 705w
The runtime difference between both systems is larger with BRP5 than BRP4. I have to look and see why the second system is running the BRP5 tasks significantly slower than my first system.
So far all completed tasks are pending validation.
I had just last night
)
I had just last night switched from Catalysts 12.1 to 13.4 on my HD6850, and I was checking if they'd run any faster, or slower than under the previous drivers.
Not much between 12.1 and 13.4 in times, on average still around 2,300 seconds. But then in came a couple of BRP5s, which run at 17,600 seconds. Ouch.
For the moment I disabled BRP5 as an option, they'll just have to wait until next month, when I'm planning to get me a HD7850 or HD7870. Probably the latter.
I'm wondering why BRP5 tasks
)
I'm wondering why BRP5 tasks have been put into queue, but their validator isn't running ?
RE: I'm wondering why BRP5
)
I'm sure there is a reason, but I'd really like to see it turned-on. I'd like to know if the work is being done successfully.
I've got about 100 of them waiting to validate, some with NVIDA some with AMD, several different OSes. If I could get one per machine to validate I'd be happy.
RE: I'm wondering why BRP5
)
- BRP5 tasks were issued because we are running short of BRP4 data (the server status page shows when the remaining data will be finished i.e. tasks returned, it doesn't show when we will run out of tasks to send out).
- Starting BRP5 was actually delayed, mainly because of me being sick most of last week.
- We started shipping BRP5 on Friday. As no one would be able to watch what goes right or wrong over the weekend, we decided to deal with validation (assimilation etc.) on Monday, when we could easily watch and possibly correct errors. Also too we'll have a better understanding of the run-time relation between BRP4 and BRP5 than what we got from the very short tests on Albert.
We're somewhat victims of our own success. The recent increase in GPU computing power led to a higher processing rate of Arecibo data than the rate at which Arecibo can capture it, and also to tasks run-times so short that the throughput began to seriously stress our DB. BRP5 is targeted to address both issues.
BM
BM
According to my short
)
According to my short observations and with assumption that the results are correct, the individual task should give about 5000 to 8000 points. I also think it would be good to have a separate option to set the GPU utilization factor for BRP4 and BRP5.
RE: I also think it would
)
That would be great!
Running two tasks in parallel works good for BRP4 tasks, but BRP5 tasks seem to run about 35-40 hours on my NVIDIA 610M.
RE: According to my short
)
Can anyone tell me what other projects (e.g. SETI) grant on average per GPU hour?
Soon there won't be any BRP4 tasks anymore for GPUs.
BM
BM