Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Path: utzoo!utgpu!water!watmath!clyde!rutgers!mit-eddie!genrad!decvax!ucbvax!sdcsvax!darrell
From: darrell@sdcsvax.UUCP
Newsgroups: comp.os.research,mod.os
Subject: Re: Life with TLB and no PT
Message-ID: <3028@sdcsvax.UCSD.EDU>
Date: Wed, 22-Apr-87 15:38:13 EST
Article-I.D.: sdcsvax.3028
Posted: Wed Apr 22 15:38:13 1987
Date-Received: Fri, 24-Apr-87 03:47:49 EST
Sender: darrell@sdcsvax.UCSD.EDU
Organization: Carnegie-Mellon University
Lines: 37
Approved: mod-os@sdcsvax.uucp
Xref: utgpu comp.os.research:6 mod.os:135

First a question: isn't this what an AMD29000 is like? It has a 32 line, 2
way set associative TLB. On a TLB miss, it takes a trap and lets you do
whatever you wish. Theoretically, traps are extremely fast. The processor
doesn't save any context for you, and it has enough registers that you can
dedicate a few of them of them to the TLB trap handler. The TLB entries are
also tagged with an 8 bit Task ID which must match the PID in the
configuration register for a TLB hit. I don't think the TLB is big enough
though. I think that TLB reload from a two level page table could be done in
15-20 cycles or so assuming all pages are in memory. I think somebody
mentioned that the MIPS chip does at least part of its TLB reload in
software, anybody care to comment?

My interest in this subject is a lightweight process model operating system
for single user workstations. Being able to design your own protection
mechanisms into the paging scheme makes doing this efficiently a lot easier.
There is also a lot of flexibility to be gained in shared memory and paging
algorithms. You aren't satisfied with the reference bits in the page table?
Go add some more...

I hadn't thought much about servers and their effects on the TLB. I am more
interested in what happens when a user has 10 different interactive processes
going at once and the mouse tracks through five of them on its way from the
editor to the shell... Of course, in such a system there should be enough
sharing that most of the same code is being called. The same applies to
server code. With lots of sharing, not that many TLB entries would be needed.
(Of the N megs of code in memory on my machine, how much of it is redundant?)

We are already starting to see useful systems for this kind of research. The
VICE filesystem operates outside of the kernel. MACH supposedly allows user
specified paging code. Perhaps this could be used to record paging activity
in detail.

Sincerely,
Zalman Stern
Information Technology Center
Carnegie Mellon University
Arpanet: zs01#@andrew.cmu.edu