The size and nature of datasets is also a performance consideration. You may choose to keep your datasets in core, on local disks, spread across a number of local disks, on H:.

The nature of the dataset determines whether you need to keep it around for handy access, whether it needs to be written in a format portable between machines, whether it is archival in nature, or whether it is temporary in nature. Binary reads and writes are faster than formatted ones.

Further considerations obtain for parallel applications. If all the nodes (for a large number of nodes) will be accessing the same dataset at once, you should not have the dataset on H:. If not all nodes need all the data from an input dataset, you could have the root node read in the data and send only what is needed to the other nodes. Even more efficient would be to use a system command ("mcp") to split up the file into ones that can be read simultaneously by each node. If datasets are written out in parallel, you need to make sure different nodes are not writing to the same dataset.

This subject is discussed in more detail in the talk on Data Management.