Xref: utzoo comp.arch:20513 comp.os.misc:1551
Path: utzoo!utgpu!news-server.csri.toronto.edu!cs.utexas.edu!swrinde!zaphod.mps.ohio-state.edu!sol.ctr.columbia.edu!emory!hubcap!mark
From: mark@hubcap.clemson.edu (Mark Smotherman)
Newsgroups: comp.arch,comp.os.misc
Subject: RS/6000 cache question; OS implications
Message-ID: <12886@hubcap.clemson.edu>
Date: 1 Feb 91 21:00:52 GMT
Organization: Clemson University, Clemson, SC
Lines: 38

I'm interested in how the cache aliasing problem was solved on
the IBM RS/6000 and any implications for the OS paging management.

The RS/6000 has a physically addressed cache in which virtual
address translation is overlapped with cache lookup.  In a straight-
forward implementation, only the invariant bits would be used for
the overlapped lookup.  That is, only the low 12 bits (specifying
the byte offset in a 4k page) would be used in lookup since they
would remain unchanged between the virtual address and the physical
address.  The page frame number, which arrives later -- after
translation -- , would be used for the tag matching.  However, this
approach limits the total cache size to the page size times the
set associativity.  (cf. i486)

However, the RS/6000 uses 20 bit lookup: the 12-bit byte offset
plus the low 8 bits from the virtual page number.  The Oehler and
Groves paper in Jan. 1990 IBM JRD has one paragraph (1st column, p.
35) on the implications of this, and the Groves and Oehler paper in
the IBM RISC System/6000 Technology book also has one paragraph
(1st column, p. 22).  However, I don't understand these paragraphs
as to what they imply for the OS.

In my simple thinking, it seems that if they use the 8 low bits of
the VPN, a page can only be mapped into a page frame whose
PFN mod 255 == VPN mod 255.  This seems to tremendously impact frame
allocation by the OS -- and cause contention for a limited number
of page frames in each mod 255 equivalence class.

(A worst case??  Lots of processes executing with locality in their
own unique virtual page 0's contending for a set of page frames
that is only 1/256 the size of main memory.)

How does the alias avoidance actually work?  How are equivalence
classes of page frames avoided?

-- 
Mark Smotherman, Comp. Sci. Dept., Clemson University, Clemson, SC 29634
INTERNET: mark@hubcap.clemson.edu    UUCP: gatech!hubcap!mark