In article <39bac689.1728385@news1.inter.nl.net>, A.Bouwmeester@inter.nl.net (Aswin) wrote: > This problem is not memory related as some suggest but I/O related. I > do not know the cause, but this is how it works: > SAS datasets are paged, each page can contain multiple obs. When > reading data from disk it is done page by page, not obs by obs. So if > you want one obs SAS reads a whole page. If you want another one that > is in the same page, there won't be any I/O. If you want a second > that's not on the same page, there is I/O. > If you use an index to access the data it is not very likely that > subsequent obs are an the same page, most obs require a new page to be > read. In the end pages are read multiple times, could be as often as > there are obs in a page. If the data was sorted you would be sure that > each page will be read only once. > Conclusion: Indexed data access can result in suboptimal I/O. > > Some hints > -Pagesize can be set using bufsize option > -Pagesize and number of observation in a page can be queried using > proc contents > > Aswin

Wouldn't BUFNO also be somewhat useful? Would setting it up to 16 or 64 or some appropriate number N force SAS to effectively cache the last N pages?

Sent via Deja.com http://www.deja.com/ Before you buy.