Re: MISC-d Digest V97 #30

[jfox@xxxxxxxx (Jeff Fox)]

Dear MISC readers:

>Why not .35u geometry?  I know from personal experience that IBM and
>Lucent are fabbing at that scale now.  Skipping over .5 would give low
>voltage/power and a speed-up as well.

I am sure when Chuck gets access to .35u he would be very happy to
use it.

>Can you describe "active messages" some more to me?  It sounds like
>what I presented as the message implementation for the Dreams OO
>extension for Forth at Rochester in 1990, and in a paper in JFAR in
>1994.

In November of 1995 I heard a great presentation by Dave Culler from
UC Berkeley about "Networks of Workstations (NOW's) Using Active
Messages" at the monthly meeting of the Parallel Processing Connection.
As I recalll they had quite a bit of info available online on the
web about their project. It involved using some small simple hardware
to route their messages and software that treated the messages in
a different way than conventional message passing implementations
built up on top of the conventional network support services.
And what follows is my recollection of one of the points he made.

When parallelism is created using conventional message passing it is
done on top of conventional network services.  These services were
designed to provide an alternative to disk access.  When one uses
these services one is using several layers of software.  A message
is processed by the network service code and stored in a queue
somewhere and a lot of code gets executed at various layers in
the network code.  Then at some later time a process that is
waiting for a message will check the whole message queue for any
messages queued up for it and if it finds any messages addressed
to it it will remove them etc.  This process can introduce latencies
that are huge.

When parallelism is created using active messages the software is
much closer to the metal.  A good deal of the conventional network
code is eliminated and when the message arrives it immediately
awakens the recipient.  The huge latencies introduced by networking
software design to compete with disks is removed.  Messages invoke
immediate action when they are sent.

The performance improvement achieved at UC Berkeley by doing this
was very impressive.  And with workstation farms (NOWs) being popular
they had a good idea for improving the bottleneck on their system.

On a parallel system you would like communications to be just as fast
between nodes as it is locally.  In such systems as SCI (Scalable
Coherent Interface) it is hidden in the hardware so the processor
sees no difference between cache access and access via the
network.  But when you have a network that is much slower than
memory you need to optimize network communication performance and
minimize latency.  Using services written to compete with disks is
not the way to go.  Using services that sit close to the network
hardware and optimizing network performance for message passing
in a parallel system is what you want.

In F21 a message DMA ends in a CPU interrupt.  If you send a message
over the network it gets put into a nodes memory with DMA and
instantly the CPU is vectored to do the appropriate thing.  The
hardware the routing DMA and CPU interrupt with only a few gates.
The software will be designed to support this function rather
than act like a disk drive.

>Concerning Beowulfs: current Pentium clones can be had for few 10 $, last

Beowulf spec calls for the fastest Pentium available.  Best case we
are still talking about an order of magnitude difference on CPU
price alone.

>time I looked MuP21 sold for roughly $40. Motherboards cost next to

The reason for MuP21h costing $40 is simply volume.  Dr. Ting had
a very small number of chips made, then he had some of those chips
packaged in PLCC at a second considerable cost.  Given his yield
I think he is losing a lot of money by only selling a few at $40.
If he sold em all for $40 he might clear a little not counting
all those years of development expenses.  If you want enough of
them they should be a couple of bucks each.

In prototype quantities F21 will cost about $500 each, and large
volume it could be $1.  So maybe we are still talking more like a
100/1 price ratio on the CPU alone.  Add the support chips, the
video, sound card, network card, extra memory, etc and then give
me an estimate that more closely matches what is actually in
that F21.

>run X, it is just needed for visualization on one machine. There's lots 
>of MPI scientific software, which can be ported to Beowulf by just one 
>"make". This code requires good float performance, which F21 can't do.
>
>Don't get me wrong, I still go for the F21. But it just doesn't compete 
>with the Beowulf, a different type of a beast entirely. (While Beowulfs 

Well it does compete on some applications that are not FP intensive.
People are just conditioned to think that parallel machines are only
for FP crunching.
 
>In that regard the SuperDSP box is like the F21 -- porting is required.

Yes, but that depends on if you use Forth!  C programs will port fairly
easily to a parallel Beowulf.  Forth programs will port fairly easily
to a parallel F21 system.  This is either an advantage or a disadvantage
depending on what kind of software you prefer to use.

>> This Saturday Skip Inskeep, John Rible, and Chuck Moore will be doing
>> presentations at the SV FIG meeting.  Skip will talk about the tethered
>
>Are planning to video tape the meeting?  I yes, I'm ready to buy a copy. 

I'll see if I can tape some of it.  The meeting goes from 10:00 AM to
4:00 PM.  Chuck is scheduled for 2:00.

>I saw that Chuck Moore will be speaking on Saturday. I was wondering 
>if anyone knew where it would be. Do you need to be a member of
>FIG? I would like to go, but need to know if this is possible.

Yes, it is at Cogswell College in Sunnyvale.  No you don't need to be
a FIG member, or an SV FIG Chapter member to attend.  You are welcome.
Take 101 or 237 to their intersection, from either highway exit to
Mathilda Ave. and take it under 237 and take the first right onto
Moffet Park Drive.  MPD runs parallel to 237 just a few feet north
of it.  Go one block (east) turn left onto Bordeaux Dr.   Go about
1/3 block to the third driveway on the left.  Park in lot and enter
main building, goto seminar rooms, or follow "Forth" signs, or ask.

>I just bought a 166 K5, which was quite a number of $10. 83$ is perfecly 
>allright considering what an F21 is going to cost ($1? $10$ $40? I think 
>the latter).

I have always said I would not like to price F21 above $21, even in
singles, but even that requires a considerable investment on my part.

>But I don't want to buy 10 k. I'd like to buy let's say 64 F21's, _with_ 
>a board, and SRAM. Besides of the fact that I can't buy them, I don't 
>think the prices will be that spectacular.

I would like to seem them priced as low as possible.  It will all depend
on interest and quantity produced.  Those are just the facts of life.
Jeff Fox 
jfox@dnai.com    Ultra Technology Inc. 
http://www.dnai.com/~jfox/