Should ERP force one case to have 2 threads?

[rljacob]

ERP is sometimes described as a combination of ERS, PET and PEM.

But for a case where the default is 1 thread, PET will force the first run to have 2 threads, so the second has 1 thread.

ERP does not do that. If the default is 1 thread, both runs will have one thread.

Should ERP be changed to force different thread counts on the 2 runs?

[billsacks]

My initial reaction is "no"... maybe better to think of ERP as a combination of ERS, PEM and optionally PET. ERP is one of our most common tests in CESM, and I think we don't want to force all of our tests to run threaded this way. However, I don't have strong feelings on that... I'm kind of curious to see how much would break if we changed this behavior....

[jgfouca]

We could always make a new test type (ERPT or something like that) if needed.

[billsacks]

Rather than introducing a new test type, I'd just encourage people to use the _P modifier in the test name to accomplish this if the default PE layout doesn't include threads.

[rljacob]

For a test-two system test, will the _P modifier only apply to the first run?

[billsacks]

For a test-two system test, will the _P modifier only apply to the first run?

The _P modifier specifies the PE layout for the first run, and then the second run will make its modifications based on this. So, for example, we have some ERP tests that have ERP_P64x2... that means we get 64x2 for the first run and then 32x1 for the second.

[rljacob]

We test on a few different machines and it can be hard to find a Tasks x Threads setting that works on all of them for a given test. Can something like "_Px2" just change the thread count and leave the tasks alone?

[billsacks]

I'm not sure if you can use "_Px2"... I don't think I've ever seen something like that.

We used to support a "_PT" layout (which would be defined in config_pes.xml)... I'm not sure if that's currently supported. See also #900

[billsacks]

I should add that I'm also open to the other ideas discussed here. For example, if it's impractical to use the _P modifier here, then I support @jgfouca 's idea of introducing another test type. My original reaction was to be worried about duplication, but if this can be done by inheriting from the ERP test and then just adding a small bit, that feels like a reasonable approach.