True...The problem may be someplace other than IO. The same large dataset is read by Proc SQL in a different job and that step takes a bit more than 2 hours to run. The SQL is a simple count with group-by on 7 fields; it returns about 1 record for every 1000 input records. I don't know if basic IO and SQL IO are radically different.

I have some stats from a run with no bufno and with bufno=100. There is no significant difference between them. I've started another run with bufno=1000 and with OBS=half-the-file. I'll post all the results this afternoon.

--Andy Arnold