<g.gorin>lieberman.memo2
Plans for LOTS, 1978-79

This is a summary of the hardware and software upgrades that we think
will adequately increase the effectiveness of the LOTS computer
system.  Further details are in the attached pages.

After a year and half of operation in service to the University, the
computer at LOTS is often found to be overloaded.  The overload is
apparent in several effects:

        1.  Lengthy delays for response from the computer
        2.  Long queues of students waiting to use terminals
        3.  Occasional long delays for printed output
        4.  Long waits to access the tape drive, and
        5.  Shortage of disk space.

The first of these conditions is most distressing. Also, each of the
other symptoms of overload can be attributed in general to
ineffectiveness of the computer system under overload conditions.

A better-running system will give better response, get people off
the terminals quicker, shorten the queue of people waiting for
terminals, and provide better printer throughput on the printers
π
we already have.  Less obviously, poor system throughput is preventing
us from taking full advantage of all the disk space we already have.
(Some further explanation is made in the attachment.)

The remedy we propose includes new hardware acquistions and some
software development at LOTS.  The new hardware that is needed
consists of:

        1. Upgrade to the 2060 central processing unit,
        2. Addition of 512 K words of memory, to double
           LOTS' main memory size,
        3. Addition of one tape drive, and
        4. No growth, or very small growth, in the number of
           terminals attached to LOTS.  No additional terminals for
           for LOTS are desired at this time.

The new software includes:

        1. Tighter limitions on the users of LOTS and the 
           extent of their use,
        2. More effective software, to allow the same learning
           experiences at a smaller expenditure of LOTS' resources, and
        3. A terminal reservation system, to allow faculty and
           students to reserve specific terminals at specific times.
π

The cost of this hardware is presently uncertain. The current "best"
price is $180,000, plus about $7,000/year in additional maintenance
charges.  However, this price includes some components that aren't
firmly priced, and involves us doing business with a vendor other
than the supplier of the mainframe. We have yet to hear the "best and
final offer" from anyone.  It will be possible for us to lease the
memory on a a short-term basis to evaluate its effectiveness.

↑LWe expect that this upgrade will increase computer throughput
adequately to serve the present level of demand from the academic
users of LOTS. With this upgrade, under "worst-case" loads we expect
the system will deliver nearly twice as much throughput (65 to 70% of
all CPU time) than it does presently (35%).

We acknowledge that making LOTS better will tempt some people to make
more extensive use of the facility. To restrict the growth of new
uses, we have limits on terminal time and the limited number of
public terminals. We may find that restrictions to limit the influx
of additional classwork may be necessary.  We don't yet know what
form such restrictions would take.

Adding to the main memory is the best technical solution to our
overloading problems that is available for implementation by Fall
π
Quarter. In the absence of increased demands by faculty and students,
these hardware additions will provide the University with an
adequately powerful system for academic computing.
↑LLOTS Computer Facility
Ralph E. Gorin, Manager
July 7, 1978

BACKGROUND:

        LOTS is operating a large DECsystem 2050 computer system to
provide interactive computing service to support instruction and
unsponsored research at Stanford University.  Of the 11,000 students
at the University nearly 3,000 use LOTS each quarter.  LOTS is used
for instruction in Computer Science, Engineering, Statistics,
Education, and the social sciences. In addition to instruction, LOTS
is also used for unsponsored research and to some extent for
word-processing by students writing dissertations and papers.

The LOTS system hardware presently includes a DECsystem-2050
processor, 512 K words of main memory, three 200-megabyte disk drives
on two channels, one tape drive, and eighty terminal ports.

Two of the three disk drives are in use for public storage and
swapping. The third disk is presently being used for large,
π
infrequently changed, data files, principally system accounting data
and program sources.

THE PROBLEM:  Demand exceeds supply.

Demand for LOTS chronically super-saturates the effective supply of
services.  We must improve the ratio of effective throughput to user
demands.  The system saturates to the point where people find that it
is no longer useful. They would quit using LOTS under those
circumstances but for external pressure from due dates.

On the demand side, we have placed restrictions on how much console
time each user is allowed per week.  We have reason to believe that
these restrictions were not entirely effective Spring quarter.
However, a better job can be done, and it will be done this Fall. We
are tightening up the account-making process this summer, to better
prepare for Fall quarter.

A recent episode shows how, despite our controls, we have failed to
reduce demand.  The experience with the Human Biology course, in
which 200 novice users, all doing resource-consuming plots, were put
onto LOTS at about the sixth week of the quarter, is one that we
don't want to repeat.  However, given that LOTS is charged with
supplying all instructional computing, we have very little means to
π
prevent a recurrence.  When Professor Durham first approached us in
this matter, it was his intention to use the computer in early May;
this looked reasonable to us. However, the special programs that he
needed weren't finished until about May 15, so the students in Human
Bio 4 contributed an extra burden to the usual end-of-quarter rush.

↑LWe hope that as the faculty gains a more thorough understanding of
the capabilities and limits of the LOTS resource such incidents will
diminish in frequency and impact.  One of LOTS' functions is to
provide for innovative pedagogical use of the computer, so we must
expect that additional courses will "discover" us. Furthermore, as
more students at Stanford become aware of the variety of services
that can be obtained at LOTS, we expect to see a gradual increase in
the services delivered outside of required classwork.

There are two things we can do to bring demand and supply into
balance. First, we can limit demand, by placing restrictions on who
can use LOTS and in what amounts.  Second, we can expand the supply.
I think that the present restrictions on demand should be more
effectively enforced and that an expansion in computer capacity is
needed.


LIMITATION OF DEMAND:
π

We propose to continue the console-time allocation system that we
started this spring.  There will be some changes to the system,
including the following:

        1. No carry-forward of unused weekly allocations.
           Computer time lost in idleness can't be reclaimed
           in the future; this rule reflects that characteristic.

        2. Time allocations to students in each course will be
           based on the computational-intensity of the particular
           course.  This way, a course such as Human Biology 4 A
           and B (a total of 8 units), wouldn't automatically get
           an allocation of computer time twice that of CS105.

        3. A tightening of the account issuing process to eliminate
           questionable accounts and poaching.

        4. Encouragment of unsponsored research users to find a
           little money and to avail themselves of SCIP's services.

        5. Enforce the per-session time limits more effectively.

        6. Establishment of a terminal reservation system, in which
π
           a terminal can be reserved for a specific time in the
           future.  This means that people can budget their time
           more effectively.

INCREASING THE SUPPLY: Effective Changes and Additions

There are two ways to increase the supply of computer resources.
First, we can buy additional computer hardware.  Second, we can
employ better software that more effectively delivers the hardware we
already have.

↑LBetter software is being tried.  However, software projects are quite
squishy in both timing and effectiveness.  We have two main thrusts
of software development underway, both with potentially large
payoffs.  But software techniques alone cannot be depended upon to
produce the drastic improvement in effective computer service that is
needed at LOTS.

        The software projects that are in progress include:

        1. The development of a new implementation of PASCAL for
        use by the introductory and intermediate Computer Science
        (and related Electrical Engineering) courses.  This project
        is being done in response to two operational difficulties
π
        that have been observed in connection with the SAIL language
        implementation that exists at LOTS.  These problems with
        SAIL, that we hope to eliminate in the new implementation of
        PASCAL, are:  Incomprehensible and (sometimes) misleading
        diagnostic error messages, which hinder students' programming
        debugging, and the use, by SAIL, of the linkage editor
        (LINK) which is a large and inefficient consumer of system
        resources.

        2. Changes to the operating to make it allocate main
        memory and other system resources more effectively.  The
        LOTS staff is too small to get into this kind of tinkering
        too deeply, but we might get lucky and find the right
        change to make.   We are also trying to influence DEC to
        make the performance improvements that we think are necessary.
        We would rather have DEC do the work because our
        limited manpower is needed for other things.


Analysis of operating data shows that in the worst cases observed,
only 35-40% of the CPU is delivered to users.  Another 20% goes to
scheduling users, and the remainder, sometimes more than 40% is
wasted due to not having anything in memory that can be run. The
problem of vast amounts of wasted time (technically known as "swap
π
wait") can be dealt with most effectively by adding more main memory.
We expect that with 512 Kw additional memory the swap wait time could
be dropped to 5-10% in the worst case, with a corresponding increase
in useful work to the 65-75% range.

PART OF THE SOLUTION: A new processor

Before going into the specifics of the proposals that have been made
to us regarding additional memory, we must first discuss the 2060 CPU
and what it means to us.  2060 is DEC's name for their top-end system
in the 20 series.  The 2060 includes a faster and more versatile CPU,
more memory than our 2050 has, and a new operating system that takes
advantage of the new features of the CPU.  Although we don't
presently need a faster CPU, the other features of the 2060 are
attractive and LOTS will find it necessary to have these features if
we are to keep up with DEC's future software developments. We think
it's better to follow DEC's operating system developments because in
the long run they will be good for us, and any development on our own
will lead to long-term high maintenance costs. It is better in this
instance to forego transitory benefits of our own development for
long-term benefits and economy of operation.

↑LTwo immediate benefits of 2060 are that it allows us to integrate
more than two disk drives into the publicly accessible disk (one of
π
our disks is presently considered private because of this
restriction) and the 2060 allows us room (technically, room in the
operating system's virtual address space) for data tables that are
needed to describe the memory that we want to add to the system.
While it is possible to rearrange the operating system so that it can
have enough space to describe all the memory and all the disks, to
use memory in this way will cut us off from future software features
that compete for this space.

I believe it is in our long-term best interests to upgrade to a 2060
processor.  All else being equal, we could postpone this upgrade
because it is not otherwise urgent.  However, if we ever upgrade to a
2060, it should be done now, because the memory that we wish to buy
will have to run with the new processor, and unless we have the new
processor, we will not sure the memory will work with the 2060
because the 2060 has tighter timing tolerances than the 2050.

The price for the upgrade to the 2060 is $50,000, including the new
processor and a version of the operating system that supports the new
processor freatures and additional memory.  [All prices mentioned
herein are exclusive of sales tax, insurance and shipping.]


THE DIFFERENCE: More memory
π

We have observed similar systems that have more main memory.  They
perform substantially better under heavy load than LOTS does.  Among
the systems we've seen are the DECsystem-1090T at SRI and the
KL10-based system at Artificial Intelligence Laboratory.  SRI has
1024 Kw of main memory, and although their system is called "1090" it
really is essentially the same hardware and software that LOTS has.
Their system runs 100 interactive users and performs much better than
ours does when we are running 60 users.  The principal difference is
that SRI has twice the memory that we've got.  The Artificial
Intelligence Laboratory has recently added 512 Kw to their 384 Kw
system.  Performance is now substantially better than it was.

There are two vendors of add-on memory.  DEC has recently announced a
solid-state memory for the DECsystem-20.  To double the existing
memory, using DEC's new memory will cost approximately $196,000 (plus
tax, shipping, insurance, and installation).  This cost breaks down
as follows:

        756 Kw of MOS memory            $250,000
        256 Kw of core, trade in        ( 54,000)
        Total:                          $196,000

The present memory is 512 Kw of core, in two 256 Kw boxes. To make
π
room for the new memory, one box has to be removed.  The trade-in
allowance on the old memory is $54,000.  The new memory will be 768
Kw, which is 256 Kw replacing the core that was removed, plus 512 Kw
additional memory. MOS memory from DEC cannot be delivered before
January 1979. MOS memory absolutely requires the upgrade to the 2060
processor.

↑LA second plan also involves DEC as the vendor, but is unorthodox.
This plan calls for us to buy 512 Kw additional core memory and
connect it to the existing system.  Engineers at DEC are uncertain
about whether such a connection would work; however, DEC is willing
to let us experiment and if the experiment reveals problems we may
return the memory.  Also, DEC has agreed to let us lease the memory
on a trial basis.  The exact details are somewhat complicated but the
proposal is roughly:

        512 K memory and cabinet,   $ 140,000
        lease of the above, $7,000/month, of which 100% is 
        applicable to purchase.  Maintenance while on lease is
        an additional charge.

It is not necessary for us to upgrade to a 2060 processor before
installing this memory.  However, there is a substantial probability
that the core memory will not work, or will work less well with the
π
2060 processor than with the 2050 we already have.  This being the
case it is prudent for us to try the extended memory configuration
with the 2060 before buying the memory.

The third plan involves another vendor, Ampex.  They propose to add
core memory to the existing system.  The price from Ampex for 512 Kw
of their core memory is $115,000.  The Ampex memory can also be had
on a short-term lease, for $5,000 monthly, with $4,480/mo applicable
to the purchase of the memory.  The disadvantages of going with the
Ampex memory are:

        1. Potential hardware problems that prevent the memory
           from working reliably.  (It's important that we
           evaluate the performance of the memory on a 2060
           system.)

        2. Difficulties in dealing with suppliers and maintenace
           in a multi-vendor environment.

        3. Gaining the enimity of our mainframe and software vendor.

Offsetting these negative considerations is the price.  Moreover,
engineers at AMPEX are confident that their memory will work, while
there is some doubt that the DEC core will work in our configuration.
π
It looks to me as though going with AMPEX is the best thing to do.

ALSO NEEDED: A second tape drive.

The present tape drive is adequate for most purposes, but when LOTS
is very busy, it is difficult for a student to use the tape drive
because it is tied up, for many hours at a time, writing our disk
backup tapes.  Because of the general slowness of the system, we
cannot effectively backup all three disk packs; this prevents us from
taking full advantage of the third disk.  It would be more than a
small convenience to have a second tape drive.  The tape drive is our
major communication path to DEC, to SRI (with whom we share software)
and to the Graduate School of Business.  Also, for our users, the
tape drive is the main path to SCIP.  We are anxious that people be
able to take advantage of both LOTS and SCIP, in whatever mix bests
suits their needs.  To not have an available tape drive inhibits
people from utilizing the best features of each system.

↑LTo some extent, having more main memory is expected to relieve some
contention for the tape drive by making the disk backups go faster.
However, if we also expand to three public disks, we'll be nearly
back in the same place as before, with respect to accessing the tape
drive.

π
Additional tape drive: $15,000.

IN SUMMARY:

The 2060 upgrade is necessary for our proper evaluation of whatever
memory we purchase.  The 2060 upgrade is necessary in the long run
to be compatible with DEC software developments and to take advantage
of new features as they become available.

2060 Upgrade:                                           $50,000

Apart from an additional entire mainframe system, nothing will help
LOTS meet the needs of the user community than more memory.  We
need to double the memory.  We have two offers that can do us some
good for Fall 1978, but the price on one offer is still squishy.

512 Kw additional memory from DEC, purchase:            $140,000
        short term lease: $7,000/month plus maintenance
        100% of the lease price applicable to purchase
        after 3 months.

512 Kw additional memory from Ampex, purchase:          $xxx,000
        short term lease: $5,000/month including
        maintenance.  4,480/month applicable to
π
        the purchase after 3 months.

Purchase of an additional tape drive:                   $ 15,000

WHAT WE WANT
        1. Funds to upgrade to the 2060 processor.
        2. Funds, up to $7,500/month for
           three months to lease a memory.
        3. A reasonable prospect that if the experiment
           with the memory is successful in providing
           greater throughput, then the memory will be
           purchased.

        4. Funds to buy an additional tape drive, $15,000.

@