[Hardware] design choices (was: language neutral processors)

[Jecel Assumpcao Jr]

Hans-Martin Mosner wrote on Thu, 02 Aug 2007 08:29:54 +0200
> When implementing Smalltalk hardware, what kind of object memory design
> would you prefer:
> - object table vs. direct pointer

The great thing about the Mushroom-style virtually addressed caches
(which Mario Wolczko is still trying to get Sun to adopt today) is that
you get the flexibility of the object table and the performance of
direct pointers. It is quite simple: you present to the cache an address
with two parts, the object ID and the offset, and if there is a hit you
get back the data directly. Only when there is a cache miss do you have
to look up in some tables to translate the object ID into a physical
address to be able to load the data into the cache.

Normally virtually addressed caches are considered a bad thing even
though they are faster than physically addressed caches. But that is due
to aliasing problems (two different cache lines might have copies of the
same data) which make life difficult for C. Since Smalltalk doesn't have
pointer math we don't have to worry about that.

http://www.wolczko.com/mushroom/index.html
http://research.sun.com/projects/dashboard.php?id=13

> - ref counting vs. generation scavenging cs -whatever-

In general I like whatever works in a distributed context. For a single
workstation generation scavenging still seems to be the state of the
art. Though these surveys are a bit old, they are still my main
reference for garbage collection:

http://www.cs.utexas.edu/users/oops/papers.html#bigsurv (but see paper
12 as well)
http://citeseer.ist.psu.edu/228113.html

> - page-based virtual memory vs. object-based VM vs. none at all

Object-based is nicer. I like not only what the Mushroom guys did but am
also a fan of LOOM and even OOZE. That said, my current design swaps
whole groups of objects (perhaps I should call them islands?) from/to
disk instead of individual objects. This is based on the fact that the
observations that led to the design of the Amoeba operating system (size
of memory vs size of files and speed of networks) are even more true
today.

> - in-band tags vs. out-of-band tags
> Let me explain the last one. Smalltalk Implementations have typically
> had in-bad tags, i.e. the tags have only meaning in a word which is
> known to be an oop. The Burroughs machines had out-of-band tags, that
> is, every memory word had an additional tag. That way, you could even
> know whether a memory word contained uninterpreted bytes, a float value,
> or an "oop" (descriptor).

My own preference would be what you called out-of-band tags. For
example, it is easy to implement a 36 bit processor in a modern FPGA
(all internal memories are a multiple of 9 bits, not 8 like earlier
chips) using ECC memory cards. The math and addresses would be 32 bits
so you would have 4 bits for tags. When dealing with Flash, disks,
networks and so on you could always compress/decompress your data and so
the mismatch between 36 and 8 bits would not be a problem.

When using development boards it isn't so easy to do this. The ML401,
for example, has 1MB of 36 bit ZBT SRAM but the 64MB of SDRAM is only 32
bits wide. And since my current focus is on making my project more
Squeak (and C) friendly, RISC42 is a 32 bit processor. I am not sure
that I would call Squeak's 1 bit tag and 31 bit data organization
"in-band", however. It certainly is when seen from the viewpoint of the
hardware or a few critical spots in the virtual machine, but for the
rest of the system it looks as much as out-of-band as the Burroughs
computers (the first real computer I ever used was a B6700, by the way)
ever were.

> Just out of curiosity: how much does one pay for such a kit? I don't
> know whether my wife would approve of a major investment into something
> of little perceived value :-)

to which Gary Fisher replied:
> The Xilinx ML401 runs about US$500.

That is exactly why I got this one - there is a limit to how much
electronics tourists can bring back to Brazil and that happens to be
$500 :-) This allowed me to avoid the absurd import taxes (which add up
to over 100%) that I paid on previous kits and the $100 Laptop.

A more modern kit like the ML501 costs twice as much but there are far
cheaper options. The two leading FPGA companies have high end families
(Virtex for Xilinx and Stratix for Altera) and low end ones (Spartan and
Cyclone respectively) and there are several smaller players with many
options. The Spartan 3E Starter Kit that Travis Kay mentioned in one of
the first messages to this list, for example, only costs $149.

> http://www.xilinx.com/xlnx/xebiz/designResources/ip_product_details.jsp?iLanguageID=1&sSecondaryNavPick=BOARDS&key=HW-SPAR3E-SK-US-G

You can get a good idea of what boards are available at:

http://www.fpga-faq.com/FPGA_Boards.shtml

I find the boards from Xess pretty interesting, though there are some
other good companies out there. The problem with these companies is that
they don't sell many boards each year, so their prices are rather high
if you only look at what components are in each board. In contrast, some
kits from the chip manufacturers are being sold below cost.

And David T. Lewis wrote:
> Well, I don't have anything worthwhile to add, but I'm certainly interested
> in your discussions. Hope you don't mind folks just lurking and enjoying
> the read :)

That is the whole point of this list. Matthew and I were having a
private discussion and realised that other people might be interested in
it. Some might have something to add while others might just want to
read this stuff. And which ones are which will probably change over
time. Having a list that is archived will allow people who come in later
(perhaps via a search) to read this.

-- Jecel