Re: A Customer's Request For Open Source Software

["Robert I. Eachus" <rieachus@attbi.com>]

Alexander Kopilovitch wrote:

> No, there is no need to fetch bunches of memory - compiler will tell (perhaps
> through OS directives) the memory chip suitable value of "base" and it's enough
> - we went through that long ago. Naturally, that will be less efficient than
> we currently associate with "real" (that is, non-virtual) memory - the switches
> of the "base" will take time, but it will be more efficient than disk I/O.

I guess this is not an Ada issue yet, but it will be. ;-)  Remember that 
when Ada first came out, there was no agreement that Integer should be 
32-bits, not 16-bits.  Of course, the only real effect on machines that 
support both 16-bit and 32-bit Integer types is on the size of Strings. 
  Within a few years, as expected, almost all implementations had moved 
to 32-bit Integers.  (Almost all, because of 24 and 36-bit 
implementations. ;-)

For a financial package is 4 Gigabytes of addressable memory enough? 
(Two Gigabytes currently under Windows.)  Some people are running up 
against that limit with spreadsheets, and most large database servers 
use machines with 64-bit address space support.  Note that with current 
large databases a 32-bit address space is insufficient to manage the 
indexes and tables that are (virtual) memory resident.

(Advertisement -- The next paragraph is in some sense a shameless plug 
for the new Opteron from AMD, and for the Athlon64 and Athlon64 FX that 
will be released next month.)

AMD is now selling processors where the architecture supports up to 
64-bit addresses.  The virtual memory system supports 40-bits of 
addressable memory per processor, and the virtual memory system supports 
up to a 48-bit address space.  (But unless you are putting together a 
system with over 1000 processors, those later numbers are details.  The 
Red Storm machine which may soon be the worlds fastest supercomputer 
will have over 10,000 Opteron processors, and I am sure they are working 
the global addressing issue.)  So not only can you now buy machines off 
the shelf that can directly address almost a petabyte of data, some 
organizations are buying them...

But what I wanted to point out about the AMD64 architecture here is that 
with a 64-bit aware OS, modules written in (x86) compatibility mode can 
call dynamically linked libraries that are written in long mode.  So a 
"database" module could address hundreds of Gigabytes of memory 
directly, while the application for the most part uses 32-bit 
addressing.  Compared to a file system call which will take thousands to 
millions of clock cycles, an in memory database can probably be accessed 
in under a hundred clock cycles.  That is a noticeable performance 
improvement.

Apple has also recently started shipping 64-bit machines based on the 
new PowerPC chip from IBM.  And if you are into masochism, there is 
Intel's Itanium2.  (Speaking only as a compiler developer, of course. 
;-)  So in a few years, almost all server hardware should allow users to 
avoid the file system and address enough memory for most accounting 
systems and databases.  Of course, someone will then invent a use for 
all that memory...

-- 
                                                    Robert I. Eachus

"As far as I'm concerned, war always means failure." -- Jacques Chirac, 
President of France
"As far as France is concerned, you're right." -- Rush Limbaugh