For OLTP on raw almost all I/O is fully aligned single block pseudo random I/O, so almost any striping will avoid hot spots and optimize concurrency, and so the stripe element size should be chosen to optimize the transfer rate, even though you do not expect to do many multiblock reads. The tip on Disk Configuration for Sequential Scans explains how to it.

My previous email assumes that the datafiles are layed ot on raw devices. I have read your tips about pseudo-random reads and pseudo-random writes, but finally when one had to decide on A stripe size what should I actually choose?

I need some tips/advice on how best to choose a stripe size (interlace size) for an OLTP database. This is required for a database re-design effort I am undertaking. Given an existing OLTP database, how best to determine the best stripe size (chunk size/interlace size), short of re-creating the database on a test box with different interlace sizes and running performance tests. As I understand, for an OLTP database, the stripe size should equal the average number of blocks written or read by an average query or transaction. This helps in limiting each average transaction to a single disk, thus maximizing concurrency. Is this correct? What other factors to take into consideration before a stripe size is decided on. Estimating the stripe size has been always one of my big mystery topics. I generally follow the guidelines of keeping it between 64K and 128K.