Network Working Group G. Hicks
Request for Comments: 392 B. Wessler
NIC: 11584 Utah
20 September 1972 Measurement of Host Costs for Transmitting Network Data
Background for the UTAH Timing Experiments
Since October 1971 we, at the University of Utah, have had very large
compute bound jobs running daily. These jobs would run for many cpu
hours to achieve partial results and used resources that may be
better obtained elsewhere. We felt that since these processes were
being treated as batch jobs, they should be run on a batch machine.
To meet the needs of these "batch" users, in March of this year, we
developed a program to use the Remote Job Service System (RJS) at
UCLA-CCN. RJS at UCLA is run on an IBM 360/91.
Some examples of these jobs were (and still are!):
(a) Algebraic simplification (using LISP and REDUCE)
(b) Applications of partial differential equation solving
(c) Waveform processing (both audio and video)
The characteristics of the jobs run on the 91 were small data decks
being submitted to RJS and massive print files being retrieved. With
one exception: The waveform processing group needed, from time to
time, to store large data files at UCLA for later processing. When
this group did their processing, they retrieved very large punch
files that were later displayed or listened to here.
When the program became operational in late march -- and started
being used as a matter of course -- users complained that the program
page faulted frequently. We restructured the program so that the
parts that were often used did not cross page boundaries.
The protocol with RJS at UCLA requires that all programs and data to
be transmitted on the data connection be blocked. This means that
we simulate records and blocks with special headers. This we found
to be another problem because of the computation and core space
involved. This computation took an appreciable amount of time and
core space we found because of our real core size that we were being
charged an excessive amount due to page faulting. The page faulting
also reduced our real-time transmission rate to the extent that we
felt a re-write of the transmitting and receiving portions of the
program was needed. In order that the program receive the best
service from the system, these portions optimized so that they each
occupied a little over half of a page. As we now had so few pages in
core at any one time, the TENEX scheduler could give the program
preference over larger working set jobs. (As an aside, because of our
limited core, we have written a small (one and one half pages) editor
in order to provide an interactive text editing service.)
The mechanism to access the network under TENEX is file oriented.
This means byte-in (BIN) and byte-out (BOUT) must be used to
communicate with another host. The basic timing of these two
instructions (in the fast mode) is 120 us per byte to get the data
onto or off of the network. A distinction was made because the
TENEX monitor must do some "bit shuffling" to ready the users bytes
to be transmitted or it must put the network messages into some form
that is convenient for the user. This is the "slow bin, bout" and
occurs once per message. If the users bytes are 36 bits long then it
will take an average of 500 us per message. If the bytes have to be
unpacked from the message to be usable, then it may take up to a
milli-second depending on the size of the message.
II. Measurements and Results
We found by timing various portions of the program that the RJS
program was using 600 to 700 us per bit byte or between 75 and 85
micro-seconds of chargeable cpu time per bit. (See tables 1 and 2 for
actual results). A short discussion of how these figures were
obtained is now in order. NOTE! We have not been trying to measure
network transmission rates; Rather, how much it costs us to take a
program (data) from our disk and send it to another host to be
Column 1 is the clock time (real-time) from when the first byte was
brought in from the disk until the last byte had gone onto the
network. (Or from the time we received the first byte from the
network until the disk file was closed).
Column 2 is computed in the same manner as column 1 except that it is
the chargeable runtime for the process.
Column 3 is the actual number of bytes that went onto or came from
the network. The letter that follows this column indicates the
direction. E.G. s for sending to UCLA, r for receiving from UCLA).
Column 4 was calculated by the following formula:
Bits per second = (real-time)/((number of bytes)*8)
Column 5 was calculated by the formula:
us/bit = (chargeable runtime)*1000/((number of bytes)*8)
Column 6 is the 5 minute load average. (See TENEX documentation for
Using these figures we can conclude that for a million bits of
information -- programs to be executed or data -- it would take 75 to
85 cpu seconds to transmit. At a cost of $474.60 per cpu hour on
TENEX's, this millionbits would cost $9.90 to 11.20 to transfer
from the originating host and potentially the same for the foreign
host to receive. This is about 33 to 37 times higher than the
predicted network transmission costs.
It is to be noticed that, in some cases, for programs to be
transmitted over the network, the cost incurred by transmitting them
was greater than the cost of executing these programs at the foreign
There may be numerous ways to reduce the cost of the network to the
(a) Treat the network not as a file device but as an interprocess
(b) Create an 'intelligent' network input/output device. This
would, of course, be customized for individual types of
operating systems and hardware configurations. For TENEX
systems this could be implemented as the ability to do mapping
operations from the users virtual memory 'directly' onto the
network. In any case, this intelligent network device would
be required to handle the various protocols for the host.
Some changes may be required in the NCP protocols.
A way to reduce the cost of the RJS program (the one measured in
tables 1 and 2) would be to change the RJS-UCLA protocol. One
possible change is to allow hosts the option of using 32 bit bytes
(because it may be more efficient!) instead of the 8 bit bytes now
required by the protocol.
Basically, it is our belief, that, in order to make the network as
viable economically as was anticipated by the authors of
reference, much work is needed on host machines and network
protocols rather than on further refinements of the communication
1. Hicks, Gregory, "Network Remote Job Entry Program--NETRJS",
Network Information Center #9632, RFC #325
2. Braden, R.T., "Interim NETRJS Specifications", Network Information
Center #7133, RFC #189, July 5,1971
3. Personal correspondence with R. S. Tomlinson of Bolt, Beranek &
Newman during the time period of 13-SEPT-71 to 19-SEPT-72.
4. Roberts, L.G., and B.D. Wessler, "Computer Network Development to
achieve resource sharing", Spring Joint Computer Conference, May
7,1970 pg 543-549.
5. Personal correspondence with Bolt, Beranek & Newman
6. Bressler, B., D. Murphy and D. Walden, "A proposed Experiment with
a Message Switching Protocol", Network Information Center #9926,
RFC #333, May 15,1972.