Classical PSHA output on different servers

[Lynne Burks]

Hello,

I am running a classical PSHA for a very large number of sites, so I parallelized the job by breaking it into many smaller regions with a subset of sites in each job. I have multiple instances of OpenQuake running on multiple servers. However, I noticed that when I run a job with the exact same inputs on different servers, the results are different. To test the sensitivity of OQ to randomness, I tried running the same job with different random seeds and investigation times on one of the servers to see if the results change. My results show no change with a change in random seed or investigation time. The only variable left is the server - is it possible that the results would change when running OQ on different servers? Have you seen this behavior before? Or do you think something else might be causing the change in hazard curve?

As an example, I've attached a hazard curve for a site in CA. The results show the hazard curve output for the same input (other than random seed and investigation time) on 3 different servers. All jobs on server 3 have the exact same hazard curve, but the jobs on server 1 and 2 are different.

Thanks,

Lynne

[Michele Simionato]

This is really strange and unheard of. Are you sure that all servers are using the same version of the engine with the same underlying libraries? Is this a UCERF calculation?

      Michele

[lbu...@oneconcern.com]

Hi Michele,

Thanks for the response! This is not a UCERF calculation, I am using inputs based on the 2008 USGS national seismic hazard map. I also confirmed that all servers are running the same version of the engine and all underlying libraries.

Do you have any other ideas on what might cause the results for a classical PSHA to be different?

Thanks,

Lynne

[Michele Simionato]

On Wednesday, February 17, 2021 at 6:23:06 PM UTC+1 lbu...@oneconcern.com wrote:

Hi Michele,

Thanks for the response! This is not a UCERF calculation, I am using inputs based on the 2008 USGS national seismic hazard map. I also confirmed that all servers are running the same version of the engine and all underlying libraries.

Do you have any other ideas on what might cause the results for a classical PSHA to be different?

No, you are the first to report this. If you are on the same machine and you change the parameter concurrent_tasks in the job.ini do you see any change?

       Michele

[Marco Pagani]

Hi Lynne, pls send us your job.ini so that we can perform a similar test on our machines
Many thanks,
Marco

--
You received this message because you are subscribed to the Google Groups "OpenQuake Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to openquake-use...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/openquake-users/0f86defd-eeee-47dd-b151-9510a6614c2bn%40googlegroups.com.

MARCO PAGANI | Seismic Hazard Team Lead | Skype mm.pagani | +39-0382-5169863

GLOBAL EARTHQUAKE MODEL | working together to assess risk

GEM - globalquakemodel.org | T - @GEMwrld | F - GEMwrld

[lbu...@oneconcern.com]

Hi Michele & Marco,

Thanks for the suggestion! I'll try changing the concurrent_tasks parameter in the job.ini and see if the results change. We currently don't set this parameter, so the default value is being used.

I will also send our input data to the engine support email so that you can run your own tests. Thank you!

Cheers,

Lynne