In a recent note, Robert A. Rosenberg said: > Date: Tue, 3 Oct 2006 15:14:28 -0400 > > >Is there any value in specifying a value other than BUFNO=1 for > >a PDSE? Why? > > I'd say YES. You want a minimum of BUFNO=2 so that as you read from > one buffer, another is being primed with data. The same applies to > writing in that there was be a buffer for you to write to as the > other is being flushed to DASD. Note that the number of buffers you > actually want is based on your processing rate. IOW: You want enough > buffers so that will always be one available while the others are > being primed/flushed. With only BUFNO=1, you will always have a delay > when you try to process past the end of the buffer. > Ummm. I'd conjecture that the CCWs never directly reference the buffers seen by the application, organized in logical blocks, but the data are always moved between those and buffers owned by the system and organized in physical pages for the actual I/O. AFAIK, the application has no control over the number of the latter used, thus no control of I/O overlap. I could be wrong; with data chaining, the application's buffers could be used directly. Would they do that, and introduce the complications of fixing pages?
For example, would BLKSIZE=16000,BUFNO=2 give better I/O overlap than BLKSIZE=32000,BUFNO=1, given that the physical I/O is always performed in 4KiB pages? Is any overhead saved by making BLKSIZE a multiple of the page size so physical pages are never split between blocks? For example, the LCM of 80 and 4096 is 20480. (Is there any overhead in each page that invalidates this computation?) -- gil -- StorageTek INFORMATION made POWERFUL ---------------------------------------------------------------------- For IBM-MAIN subscribe / signoff / archive access instructions, send email to [EMAIL PROTECTED] with the message: GET IBM-MAIN INFO Search the archives at http://bama.ua.edu/archives/ibm-main.html
