Path: utzoo!mnetor!uunet!seismo!esosun!ucsdhub!sdcsvax!darrell
From: lamaster@ames.arc.nasa.gov (Hugh LaMaster)
Newsgroups: comp.os.research
Subject: Re: LRU/NUR for TLB ?
Message-ID: <4678@sdcsvax.UCSD.EDU>
Date: 1 Mar 88 02:12:40 GMT
Sender: darrell@sdcsvax.UCSD.EDU
Organization: NASA Ames Research Center, Moffett Field, Calif.
Lines: 19
Approved: mod-os@sdcsvax.uucp

[ Hugh:  what are the interesting things about the OS for the Cyber 205? ]
[ I have seen NOS and was not impressed; has CDC learned how to write an ]
[ OS after all these years?  --DL                                        ]

The Cyber 205 actually does full LRU in hardware for the entire page table
(including the part that is in the 16 entry TLB).  Being a memory to memory
vector computer, it doesn't cost much.  However, it does have drawbacks.  The
page table is global, and the hardware does global LRU.  In my opinion,
LRU should be by virtual memory segment.  On the 205, special non-page
table entries have to be made to manage process working sets individually,
which is actually what you want done in the first place.  Shared memory
complicates things further- whose working set do those pages belong to?
If you are not careful, shared memory pages which are frequently used by
many processes can force everything else out of memory, and increase the 
paging rate. 

Query:  Are there any operating systems out there which intelligently
handle process working sets in the face of multiple shared virtual memory
segments?