When io_layout is set to 1,1 - the output becomes serialised and a bottleneck for ESM1.6 throughput. Changing to parallel io with io_layout of say 8,5 removes that bottleneck (and improves ESM1.6 throughput by 10%); however, then results in (rare/somewhat frequent) collation failures. Opening the issue here to make the connection to the underlying model - this is the tracking issue in ESM1.6-configs.
Long-term option could be investigate how to properly switch to PIO (where there is previous work done by Marshall and Rui).
When io_layout is set to
1,1- the output becomes serialised and a bottleneck for ESM1.6 throughput. Changing to parallel io withio_layoutof say8,5removes that bottleneck (and improves ESM1.6 throughput by 10%); however, then results in (rare/somewhat frequent) collation failures. Opening the issue here to make the connection to the underlying model - this is the tracking issue in ESM1.6-configs.Long-term option could be investigate how to properly switch to PIO (where there is previous work done by Marshall and Rui).