Hi Adios,

I'd appreciate some suggestions on properly configuring BP5 for our application, especially to reduce memory consumption:

• We use adias2 as file I/O, so no stream, interaction, code coupling is needed.
• It's a mpi program involving 20k cores or even more
• Every some (say, 5000) steps we write, say, ten, 1d arrays of size ~10^11 or bigger to a file.
• Every, say, 1000 steps we write a 3d array of size (20, 1e5, 1e5) or even bigger to a file.
• right now, for all io we set the substream number to 0, hoping to have one data file per node.

We are happy with the I/o speed so far, but in some extreme cases, the first (1d-array) writing fails due to out-of-memory errors.

Could you advise on strategies like number of substreams, aggregators, buffer size, shared memory size, etc., based on the machine memory and core number per node.

I understand my description right now might be vague but I can add more details to help with your advice.

Thank you again for making this great piece of software.