Network Working Group                                               Raj Kanodia
Request for Comments: 662                                      Project MAC, MIT
NIC #31386                                                            Nov. 1974



       PERFORMANCE IMPROVEMENT IN ARPANET FILE TRANSFERS FROM MULTICS




With the growth of Multics usage in the ARPA Network community, users often
need to transfer large amounts of data from Multics to other systems. However,
Multics performance in this area has been less than optimal and there clearly
exists a need to improve the situation. Even within Project MAC there are users
who regularly ship files back and forth between Multics and other Project MAC
computer systems. Recently the Cambridge Information Systems Laboratory (CISL)
development Multics system was connected to the ARPA Network. This has
generated considerable interest among the members of the CSR (Computer Systems
Research) and CISL groups in using the Multics file transfer facility for tran-
smission of Multics system tapes (MST) from one Multics system to the other.
At currently realizable data transfer rates, transmission of a typical MST,
(approximately 12 million bits), takes about half an hour. The usefulness of
the present file transfer system is severely limited by its low bandwidth.


One of the reasons for such a poor performance is the output buffering strategy
in the Multics IMP DIM (IMP Device Interface Manager.) With the hope of making
a significant performance improvement, we recently changed the IMP DIM buffer-
ing strategy. To assess the effects of this change an experiment was conducted
between the service machine (MIT Multics) and the development machine (CISL
Multics.) This memo reports the experiment and its results.


PROBLEM

Due to reasons that are of historical interest only, the output buffers in the
IMP DIM have been very small; each buffer can accommodate up to 940 bits only.
With the addition of a 72 bit long header, the resulting message has a maximum
length of 1012 bits, which in the Network terminology is a single packet
message. Due to this buffer size limit, Multics can transmit single packet
messages only, even though the network can accept up to 8096 bit long messages.
This results in increased overhead in the set up time for transmission of large
number of bits.


An old version of the IMP-to-Host protocol requires that a host may not transmit
another message on a network connection unless a Request-for-Next-Message (RFNM)
has been received in response to the previous message. Even though this
restriction has now been relaxed, the protocol does not specify any way to
recover from transmission errors that occur while more than one RFNM is pending
on the same connection. If the constraint of transmitting only one message at a
time is observed there is at least some potential for recovery in case of an
error. 8eing reliability conscious, Multics observes the constraint imposed by
the old protocol. Following the old protocol introduces delays in the
transmission of successive messages on the same link and therefore the overall
bandwidth per connection is reduced.

                                      -1-

SOLUTION


At least one method of improving the file transfer bandwidth is obvious.
Increasing the size of IMP DIM output buffers will allow us to transmit
significantly larger messages<s1>s This will result in fewer RFNMs and delays and
improved bandwidth. We have made two assumptions here. The first assumption is
that the delay introduced by RFNMs does not increase with the size of the
message.<s2>s Our experience with the network tends to support this assumption.
The second assumption is that actual time taken to transmit a message increases,
at best, linearly with the size of message. This second assumption does not
hold for a heavily loaded network. But for our purpose we may assume it to be
essentially correct.


There is another advantage in transmitting large messages. For a given amount
of data, fewer messages will be transmitted, fewer RFNMs will be read, and
therefore time spent in the channel driver programs will be reduced. Since the
channel sits idle during at least some of this time, the idle time for the
channel will be reduced, resulting in improved channel bandwidth.


EXPERIMENT


We changed the IMP DIM to implement two kinds of output buffers. One kind of
output buffer is short and can hold single packet messages only. The other kind
of buffer is, naturally enough, large and can hold the largest message allowed
by the network. If the number of bits to be transmitted is low (this is mostly
the situation with interactive users,) a small buffer is chosen and if it is
large (for example, file transfers,) a large buffer is chosen.


The new IMP DIM was used in an experimental Multics system on the development
machine at CISL. The service machine at MIT had the old version. Large files
were transmitted from the development system to the service system and the file
transfer rate was measured in bits per second ( of real time.) To get an
estimate of gain in the performance, files were transmitted from the service


<s1>s In one situation it may not be possible to transmit large messages. The
number of messages and bits that a host can transmit on a particular connection
must not exceed the message and bit allocation provided by the receiving host.
If a receiving host gives very small bit allocation, then the sending host can
not transmit very large messages. Since Multics always gives out very large
allocations, there should be no problem in Multics to Multics file transfers.

<s2>s It should be noted that the size of RFNM is fixed and does not change
with the size of message.







                                      -2-

system to the development system and the old file transfer rate was measured.
The same file, approximately 2.5 million bits long, was used in both
experiments. The BYTE-SIZE and MODE parameters of the ARPANET File Transfer
protocol were set to 36 and "image" mode respectively. This implies that no
character conversion was performed in the file transfer. The following table
shows the results obtained:

Service to Development Development to Service
(old system) (new system)

average file-
transfer rate 6,837 27,578
(bits per second)



The following information regarding the environment of the experiment is
supplied for the sake of completeness. At the time of this experiment, the
service system was lightly loaded (the system load varied between 30.0 and
35.0). The service system configuration was 2 cpu's and 384 K primary memory.
The development machine configuration was 1 cpu and 256 K of primary memory.
The development system load was limited to the file transfer processes and the
initializer process. The MIT Multics is connected to the IMP number 44 (port 0)
by a rather short cable (approximately 100 feet long.) The CISL Multics is
connected to the IMP number 6 (port 0) by an approximately l5OO feet long cable.
8oth IMPs are in close physical proximity (approximately 2000 feet,) and are
connected to each other by a 5O kilobits per second line. The results given
above show considerable improvement in the performance with the new IMP DIM.


Since there is considerable interest in the Multics development community in
using the file transfer facility for transmitting Multics System Tapes between
the two systems, we are providing here an estimate of time that would be taken
to transmit the current MST. MST 23.4-0A contains 334,231 words. When
multiplied by 36 (the word length in bits) this yields the total number of bits
to be approximately 12.5 million. Assuming a file transfer rate of 27,500 bits
per second, it will take approximately 7 minutes and 30 seconds to transmit the
system 23.4-0A.


In the experiment outlined above, there was only one file being transferred at
any given tine between the two systems. We conducted another experiment to get
an estimate of performance in the situation of multiple file transfers occurring
simultaneously. This experiment consisted of first two, and then three,
simultaneous file-transfers from the development system to the service system.
These file transfers were started by different processes logged in from separate
consoles. Because these file-transfers were started manually, we could not
obtain perfect simultaneity and results obtained for the total bandwidth are
essentially approximate. For two simultaneous file transfers the total bit rate
was approximately 40,000 bits per second and bit rate for each file transfer was
approximately 20,000 bits per second. For three simultaneous file transfers,
total bit rate remained at approximately 40,000 bits per second and bit rate for
individual transfers was approximately 13,500 bits per second.




                                      -3-