Discussion Thread for the Continuous GW Search known as O2MD1 (now O2MDF - GPUs only)
Language
Copyright © 2024 Einstein@Home. All rights reserved.
Copyright © 2024 Einstein@Home. All rights reserved.
The "error while computing" I
)
The "error while computing" I had, all ended with an "ENOMEM" error, which I interpreted as out of memory error. So I went back to running only 2 tasks in parallell, after that there were no more "error while computing". Browsing through your list I found one linux host with that error and it reported "Memory allocation error". The windows hosts put out many rows of log text but no description of the error.
I am also seeing a few "validation error" units, some of them had many attempts at them, up to eight fails.
The lone rider's attempt was also rejected btw. and the 736 unit is now queued up at two new victims.
floyd wrote:At least part of
)
Thanks for pointing that out. I had'nt noticed higher memory usage of these tasks. I looked and saw 1400 and 1700 MB for two of the GTX 960's. One was being pushed with 2020MB at that moment. These are 2GB models. It might well be that lack of memory had been causing problems at some moment.
Is there a CPU equivalence of
)
Is there a CPU equivalence of the current VelaJr1 app? I am curious how long it runs. It takes around 30 min on my GTX 1080Ti. GTX 1080Ti consumes double the power of my Ryzen 2700X which has 16 threads. This kinda means unless this app is at least 32 times faster than the CPU equivalent, it's actually more power efficient to run it on CPU. If that's the case, I probably should switch to the CPU app and reserve GPU for something else.
wujj123456 wrote:Is there a
)
Sure. https://einsteinathome.org/apps.php
At least the Beta version 2.05 sends Vela tasks, I haven't tried regular 2.00.
I run a few of these last night. 4 tasks concurrently in about 42000 sec. Host has old Xeon 6c/12t @ 3.8GHz, comparable with old i7. Your Ryzen would be faster than that for sure, but I don't know how much time per task would slow down if you run 16 concurrent tasks. Try out, it would be interesting to see.
Richie wrote:wujj123456
)
Thanks. 42000/1800 is 23 times of speed up, which means power efficiency wise GPU might actually be losing, unless 2700X runs much longer than 42000s per WU when fully loaded. Also 2700X isn't last-gen already...
How did you get 2.05 tasks? Other than "Run test applications", which I set to "yes", I don't see beta specifically called out in the list of applications I can pick. Anyway, I enabled CPU for e@h and will wait and see if/what tasks I get. Even if I don't get all 16 threads filled with the tasks I want, the other cores/threads will still be crunching other tasks/projects so the results might still be useful.
I seem to be running the Vela
)
I seem to be running the Vela cpu tasks in the 38K-40K range on my 2700X with the 2.04/2.05 apps. I ran the 2.00 app in the 43K range for the Vela tasks.
https://einsteinathome.org/host/12600970/tasks/4/53?sort=desc&order=Task+ID
https://einsteinathome.org/host/12600970/tasks/2/53?sort=desc&order=Task+ID
I also run the gpu GW app concurrently with the cpu app.
https://einsteinathome.org/host/12600970/tasks/2/54
wujj123456 wrote:How did you
)
I set it just like that. Should be enough to get them.
I actually believe tasks will then run for more than 15 hours if you run 16 tasks parallel. Maybe it would be a good idea to leave one or two threads free. That won't change the whole picture much at all, but it could give some 'air to breathe' for the system. v2.05 CPU tasks on my computer needed 670MB per task. Your host has 16GB of RAM, which should be enough even for max scenario, but these are quite heavy to run. I recall seeing some results with Ryzen but I don't remember exactly the core amounts and run times. Also the tasks were some older type as it was last week or so. But my feeling is that your GTX 1080Ti will be a winner against that Ryzen.
What's the configuration like... how many threads for CPU and GPU ?
Nice. I had a flaw in my
)
Nice. I had a flaw in my previous reasoning. While 1080 Ti's TDP is 250W, it's only using around 60% when e@h is running, reported by both HWMonitor and MSI Afterburner (probably querying same API). That's 150W. Meanwhile, HWMonitor reports around 125W package power for my 2700X under full load. That's inline with the reviews I see.
On my 1080Ti, the p50 is 1900s per WU (1700s - 2300s). https://einsteinathome.org/account/tasks/4/54?page=4&sort=desc&order=Run%20time
If I assume our setup is similar (I am using default, 100% cores and each e@h task is 0.9 CPU + 1 GPU), with 39000s on 2700X, 1080Ti is 20x faster than one core of 2700X and 25% for the whole CPU. 150W is 20% more than 125W. So they are pretty close in power efficiency, and both are last-gen hardware. Quite a far cry from the previous e@h apps where GPU pretty much decimates CPU in power efficiency, but I already read somewhere the devs said the new workload is not as good on GPU.
Which AMD GPU generation or
)
Which AMD GPU generation or model is currently recommended for GW GPU work? (Linux)
I think the Radeon VII is
)
I think the Radeon VII is still the king o the heap because of its memory and 1/4 FP64 rating.
Radeon VII | FP32 = 13.4 TFlops | FP64 = 3.36 TFlops (1:4) | mem bandwidth = 1024 GB/s
Next up would be the RX580
RX 580 | FP32 = 6.175 TFlops | FP64 = 385 GFlops (1:16) | mem bandwidth = 256 GB/s