What ever happened to Tandem and NonStop OS 2091

When I did the FreeBSD VM mgmt algorithms (upgraded from the original MACH stuff), I made sure that even a low memory system will do a good job of converging to a reasonable working set. (This is a little off topic, but still generally applicable.)

The key is to maintain a memory of page usage that is longer than the effective clock cycle, making sure that the I-O pending queues don't get so long as to overshoot the cleaning and freeing needs, and various other prudent choices.

One of my goals was (on an early version of FreeBSD) was for X to be able to start in 4MB, and eventually settle down to a reasonable working set reasonably quickly. This also included a usable (but somewhat painful X term.) Other OSes of equivalent clbutt of that timeframe did tend to continue to thrash and TENDED to converge much more slowly to a reasonable working set. At the time, my tests of FreeBSD with X windows did work (poorly) with less than 4MB.

This 'advantage' for FreeBSD is probably of less use than it might have been when I did the code, and would probably have been of much more use perhaps 5-10yrs before I wrote it. :-).

Any one of the FreeBSD enhancements really didn't make a huge difference, but all of the changes (e.g. short queue lengths, etc) did make a significant difference.

One example of the 'improvement' could be thought of as avoiding overshoot in the pageout daemon cleaning operation. If the system would overshoot the cleaning-freeing operation, then alot of pages would be unnecessarily freed and make the system less performant. I tried the same kind of stats methods for the buffer cache as for the general page memory, and the benefit was significantly less apparent, but sometimes noticeable. As time progressed, the buffer cache mgmt algorithm wasn't retained, but the paging algorithms were indeed carefully maintained.

John