Re: extended addressing

[Robert W. Kerns <rwk at MIT-MC>]

Sorry to take so long to getting around to reading your message.
It sort of got buried in the depths of my mail file for a while.
Anyway, you seem to not fully understand the VAX architecture in
comparing the address space of the VAX with that of the 10.

     
     By the way, I am not clear why people think that VAX has a
     larger	; address space than the 20 (or do people think that?) As far
     as the architecture, the 20 has 30 bits and the VAX 31 (one bit is
     used by the system).  But clearly 30 bits of 36-bit words are more
     than 31 bits of 8-bit bytes.  Now the obvious reply is that the
     existing 20 monitor does not support more than about 23 bits at the
     moment.  However I have done the following calculation to see how much
     one can actually use on the VAX.  The problem turns out to be page
     tables.  The VAX has a 512 byte page size.  Suppose that one 32-bit
     word is required in a page table for each page.  This means you need
     almost 10% as much physical space for the page table as your virtual
     memory (4 bytes per 512 bytes).  So with a 3 Mbyte machine you can get
     only about 30Mbyte of virtual memory before filling up all of physical
     memory with page tables.  Of course something is likely to give out
     before that.  But suppose that you can really get 32Mbyte of virtual
     memory.  That is in fact slightly less than the 23 bits that are
     implemented on the 20 (assuming 2 words per cons cell on the 20 and 8
     bytes on the VAX).  It seems to me that 512 bytes is a strange page
     size for a machine intended to have 31 bits of virtual address space.

First, you make a numerical error.  One 4-byte PTE for each page is
the right number, but that only works out to 4/512 or 1/128, or almost
1%, not 10%.  Second, Unlike on the -20, VAX page tables are not wired
into memory, but can themselves be paged.  This may sound strangely
recursive, but rest assured that it ends at one level; the page tables
for the page tables are wired.  Thus physical memory is consumed at one
longword (32 bits) for every 128 pages, for a ratio of 2^14 to 1.

Now, 32 Mbytes is sort of a random number, and I'm not sure where it
comes from.  I believe it may be a former or even current limit on
what VMS will support.  But to compare with the -20:  2^23 36-bit
words (8 MQ) or 2^30 36-bit words (1024 MQ) vs. 2^29 32-bit words,
or 2^28 if you don't use P1 space (normally used for the stack, etc.,
and allocated DOWNWARD).

Now, lets consider the real problem.  For each user you have this big,
you need a large disk drive.  On the -20, for example, a full 30-bit
address space would require something like a 5 Gigabyte disk drive.
(or smaller ones totalling that).  It seems clear to me that for the
current time, the limit to address space in both the VAX and the -20,
is not how many bits the processor uses, but rather how much disk space
you can afford for paging.  Yes, folks, we're limited by the physical
memory size again...