Why am I getting so many Invalid results?
3 Sep 2019 17:59:21 UTC
Topic 219517
(moderation:
I keep getting more invalid then valid results on Einstein. I run 2 other projects and all seems fine.
My computer and GPU seem to be ok, to the best of my knowledge.
Any insight or help would be appreciated.
Thanks,
Ace
The short answer is that the
)
The short answer is that the project has been testing gravity waves on GPUs and the test seems to have been recently suspended due to poor results. I looked at your results and all but one invalid were gravity waves. Your computer is probably OK.
Thanks for your
)
Thanks for your response.
Here is another question on this topic. If my computer is giving an invalid why are the other 2 people validating the same result of that particular Work Unit not getting invalids...or is it just something random?
Ace Casino wrote:Thanks for
)
It means all three of the quorum members got past the initial sanity checks considering their work alone.
Then a comparison of results was made by the validator. Had the initial comparison been "close enough" both of you would have been declared valid, and no third task would have been sent out. But they were not close enough, so a tie-breaker unit was sent out. When it came back, it turned out to agree with the other one that was not from you, so those two validated. Yours is the odd man out, so invalid.
Thanks Archea86. I know
)
Thanks Archea86.
I know how the process works. My question was really, if I’m getting invalids on a particular work unit, How are these other 2 people getting valid’s if the gravitational WU’s are messed up, or whatever the issue is?
AFAIK, the GPU wu are being
)
AFAIK, the GPU wu are being validated against a CPU. If it goes inconclusive the tie breaker is sent to another CPU. The 2 CPUs compare quite nicely and they validate and your GPU is declared invalid.
This attempt to get GWs to work on GPUs was a noble attempt and good science but it is a work in progress so we will crunch pulsars on our GPUs for a while and watch our RAC shoot up.
Ace Casino wrote:...How are
)
It's nothing to do with tasks being "messed up". Where does that idea come from? It's entirely to do with the fact that there are 2 different applications - a long standing and trusted app running on a CPU core and a new application designed to perform parts of the calculations (hopefully much faster) by taking advantage of the parallel processing capabilities of a modern GPU.
The GPU app is under development and the testing of it is revealing that it's not always getting close enough answers to those of the trusted CPU app, in certain circumstances. By analyzing the reasons for the discrepancies in the results, the Devs hope to achieve a GPU app that will both save time and mostly always get close enough to the same results. It's an ongoing process and the invalid results should hopefully lead to newer versions of the app that can achieve the required computational accuracy.
Cheers,
Gary.
Ok, thanks for all the
)
Ok, thanks for all the answers and help.
I'm getting a small but
)
I'm getting a small but consistent number of invalids, about 1.5% of all tasks, across three different machines. The only thing I can think of right now they have in common as that they all have multiple GPUs? Or perhaps a driver issue? I have assigned a full CPU core to each task. One of them is running two tasks per card. Posting here in case it's something obvious before I go around messing with the drivers (two are Ubuntu machines and driver updates can be irksome...).
Examples from each machine:
https://einsteinathome.org/task/886195352
https://einsteinathome.org/task/886217130
https://einsteinathome.org/task/886282878
Azmodes wrote:I'm getting a
)
You're posting in a thread where invalid results for the gravitational wave GPU app are being discussed. There has been an ongoing issue with high rates of invalid results for many people for the GW search. A new app has just been announced.
Your results are for a completely different search - the long running and stable gamma-ray pulsar search (FGRPB1G) where a 1-2% rate for invalid results is pretty much par for the course. Validation is pretty strict and different GPU models and different OS and drivers may give slightly different results, sufficient to upset the validator occasionally. You shouldn't interpret this as something that indicates a particular problem that needs fixing. It's the same for everybody.
Cheers,
Gary.
Ah, didn't want to create a
)
Ah, didn't want to create a new thread and the title/OP didn't suggest it being about a specific app only. My apologies and thanks for the quick and helpful response.