It's been reported within the EC-Earth consortium with their version of IFS, and confirmed with OpenIFS 40r1, that occasional random bit-flips can occur when the model writes restart files. This means that two identical runs, which are restarted multiple times, may not give the same results as each other. Note this does not affect a model run that does not need a restart, regardless of whether or not the model writes restart files as it runs.
This has been logged as a bug on our internal OpenIFS issue tracking. A fix is not likely soon, but when there is more information it will be posted. here.
For those with access to the EC-Earth development portal, this is issue #631. The OpenIFS internal issue is OIFS-305 - Getting issue details... STATUS .
Glenn
2 Comments
Etienne Tourigny
Hi Glenn, I assume this issue is still present with oifsr43r3? Do you happen know if it is still present in IFS CY47R1 ?
Glenn Carver
Hi Etienne,
This is something that Uwe I think, found when testing OIFS 40r1 in EC-Earth. It was not clear exactly what the problem was but seemed to be related to writing GRIB data via the eccodes library, though that's not clear. I have not looked at it personally and in honesty I doubt I will have time any time soon. In any case, the plan with EC-Earth is to handover all model output to the XIOS I/O server and output directly to netcdf. This will need testing again to see if the same problem occurs.
It'll be interesting to see if it does because that will point to a problem with the model rather than the output I/O. Might be worth flagging this on the EC-Earth portal, issue '631'.
Glenn