[ih] internet-history Digest, Vol 62, Issue 19

[Stephen Casner]

Probably in the more-than-you-cared-to-know department regarding Dave
Retz and ELF, some comments in-line below from my personal experience
to add to what Jake, Dave and John have said...

On Tue, 28 Aug 2012, Elizabeth Feinler wrote:
> > From: Dave Crocker <dhc2 at dcrocker.net>
> > Cycling back to the topic of Arpa's wanting to share resources, when we
> > tried to get funding for 32K more memory for the system, Arpa instead
> > said we should get a PDP-11 and use one of the terminal concentrator
> > systems (ANTS from Illinois or ELF from Santa Barbara) and do remote
> > computing.
>
> I believe that ELF was done by Dave Retz while he was at SRI.  Most
> DEC PDP-11s used it until it was superseded by UNIX.  I think there
> is an ELF manual (or some documentation - not sure) at CHM.

On Wed, 29 Aug 2012, Dave Crocker wrote:
> I believe Dave was at UCSB.

According to [1], Dave Retz was a graduate student at UCSB around
1966, and he moved to SRI in 1974.  For at least part of 1974, Dave
Retz was at Speech Communications Research Lab (SCRL), located in a
beautiful hilltop location above Santa Barbara, and that is where a
good portion of the ELF work was done.  I recall my first business
trip after joining ISI was when a few of us from ISI went to SCRL for
an IETF-style code sprint to work on ELF.  The reason for my
involvement was that ELF was to be the OS for the packet speech
project as well.

On Wed, 29 Aug 2012, John Day wrote:
> Yes, Retz was at UCSB.
>
> ANTS (I and II) and ELF had the same basic OS architecture.  ANTS was written
> in a high level language and ELF in PDP-11 assembler.  They got basically the
> same throughput, (which was considered inadequate at the time).  (300
> messages/sec sticks in my mind but don't quote me.)
>
> There was a paper at conference about 1975 in which a group at Waterloo tried
> to solve the same problem with the same OS structure and got basically the
> same throughput.  The Waterloo group had no contact with the ARPANET efforts,
> so I concluded that what we were seeing were the limits of the architecture,
> not the programming.

At ISI, we found that ELF's performance was not quite up to what we
needed for the packet speech work.  In particular, the context switch
time was too long and the rate at which data was moved between the
user and kernel address spaces in the PDP11/45 (for example, in
network I/O) was too low.  We decided to develop our own OS as a
replacement.  It was called EPOS, for Environment for Processing of
Online Speech, and first became operational in mid-1975.  Like ELF, it
was implemented in assembly language.  We achieved a context-switch
time of 125 microseconds and improved the inter-address-space transfer
rate by something like a factor of four.

EPOS was installed at a number of ARPANET sites, not all involved in
the packet speech project.  It was in service for 10 years, including
as the OS supporting the ST Gateway developed by Lincoln Lab for use
at the Wideband Satellite Network sites.

> Today's OSs mainly use the same architecture we had for ANTS, ELF, and UNIX.
> They have solved performance issues by using Moore's Law. IOW, they didn't.
> ;-)
>
> The problem as we saw it was to get speed while maintaining the properties of
> modularity and good system design.  Most of the overhead was in context
> switching and data copying.  It was clear that just optimizing code, i.e. do
> the whole thing in assembler was not the answer.

Indeed, those are the limitations, but the level of optimization
available to us was at least sufficient to solve our problems at the
time.

                                                        -- Steve

[1] Robert M. Gray, Linear Predictive Coding and the Internet
Protocol, now Publishers, 2010.