RTTOV parallel capability

[Kihang Youn]

Hi RTTOV developers,

I am trying to optimize and fully utilize the RTTOV performence by using the subroutine “rttov_parallel_direct”.

As far as I understand, the OpenMP parallelization is applied at the loop of iband as shown below.

!$OMP PARALLEL DO PRIVATE(iband) NUM_THREADS(mthreads) SCHEDULE(DYNAMIC)
DO iband = 1, nbands

Since my program has only 10 bands (i.e. nbands=10), so I can use only 10 OpenMP threads for performance improvement(even I have 76 cores though).

At this step, Are there any workarounds or techniques that would improve the performance rather than what I’m getting.

Thank you for your attention.

[James Hocking]

Hi,

RTTOV allows you to simulate multiple profiles at once, so you could pass 7 or 8 profiles into RTTOV in each call (for example), which means that for simulating 10 channels you could use 70 or 80 threads respectively. You might find you get even better performance by passing even more profiles in per call.

Best wishes,
James

[Kihang Youn]

Hi,

That’s great! That seems to give better performance.

To be honest, I don’t know how to adjust the profile file or nprof because I’m receiving an optimization task for RTTOV from other researchers. So I think it will take time because I discuss the idea with the researcher and proceed.

As soon as I get new performance results, I will share them with you.

Thank you,
Kihang

[Author]

Posts