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[1] 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[2].  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



Hicks & Wessler                                                 [Page 1]

RFC 392        Measurement for Transmitting Network Data  September 1972


   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[3].  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[3].

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
   executed (processed).

   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)




Hicks & Wessler                                                 [Page 2]

RFC 392        Measurement for Transmitting Network Data  September 1972


   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
   details.)

   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[5], 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[4].

   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
   host!

III.  Analysis

   There may be numerous ways to reduce the cost of the network to the
   host:

      (a) Treat the network not as a file device but as an interprocess
          communications device[6].

      (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[4], much work is needed on host machines and network
   protocols rather than on further refinements of the communication
   devices involved.




Hicks & Wessler                                                 [Page 3]

RFC 392        Measurement for Transmitting Network Data  September 1972


References

   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.































Hicks & Wessler                                                 [Page 4]

RFC 392        Measurement for Transmitting Network Data  September 1972


Utah-10 Accounting for Network Usage

   for the period 16-SEP-72 12:48:34, ending 19-SEP-72 13:56:11

   Clk Tim   Cpu Tim   # of Bytes     Bits/sec     us/bit   Load

        14    11.61       18930 s     10152.175     76.67    3.74
     02:56    37.89       59066 r      2670.857     80.20    3.51
     02:18    22.71       35377 r      2038.682     80.24    2.98
     01:31    34.37       56608 s      4966.431     75.89    3.35
        13    11.57       19094 s     10985.401     75.72    4.07
     04:03    42.03       63067 r      2069.297     83.30    4.95
        03     1.82        2906 s      5932.126     78.37    5.58
        45    23.58       35505 r      6237.976     83.00    5.37
        09     2.08        3243 s      2804.757     80.21    3.60
     03:28    39.25       58632 r      2246.727     83.69    4.86
        05     4.60        7470 s     10192.734     76.99    1.12
        23    10.83       16525 r      5565.378     81.95    1.17
        06     4.32        7142 s      9116.962     75.64    1.44
        14     8.56       13223 r      7170.338     80.95    1.29
        11     4.42        7142 s      4795.300     77.43    1.89
     01:34   13.287       19562 r      1659.819     84.86    2.50
        37    10.35       16183 r      3439.807     79.97    3.02
     02:43    34.49       56444 s      2764.170     76.38    3.74
        38    10.51       16196 r      3400.467     81.13    0.69
        45    34.12       56280 s      9820.704     75.75    2.57
     03:46    36.09       56280 s      1990.601     80.16    4.06
        11     2.75        4085 r      2774.900     84.30    4.86
        15     2.88        4085 r      2154.252     88.07    4.86
     01:54    11.40       16125 r      1124.203     88.39    5.12
     01:14    35.10       56280 s      6057.068     77.96    6.10
     01:07    10.67       16125 r      1919.986     82.70    1.89
     04:28    36.32       56362 s      1679.377     80.56    5.52
     02:12    17.71       27120 r      1634.818     81.62    1.73
     06:59    41.88       64333 s      1226.980     81.37    6.66
        37     7.63       12082 r      2552.243     78.97    0.64

Utah-10 Accounting for Network Usage

   for the period 13-SEP-72 2:23:12, ending  16-SEP-72  11:47:07

   Clk Tim   Cpu Tim   # of Bytes     Bits/sec     us/bit    Load

      10       2.09        3079 s     2343.227      84.77    3.80
   11:09     138.20      204596 s     2444.733      84.43    3.68
   06:16      34.78       49994 r     1062.961      86.96    3.95
   01:57      16.25       24971 r     1693.451      81.34    2.92
   12:07     114.70      183598 s     2019.577      78.09    6.79



Hicks & Wessler                                                 [Page 5]

RFC 392        Measurement for Transmitting Network Data  September 1972


   01:13       0.92         845 r       91.683     135.80    2.12
      05       5.07        7373 s    10842.647      85.99    1.93
   03:09      42.10       62414 r     2633.655      84.31    3.86
   13:22     115.13      183352 s     1828.467      78.49    0.58
      02       0.25         233 s      907.056     134.12    6.05
   07:10      44.23       64869 r     1206.001      85.23    5.07
      04       0.33         233 s      402.679     179.18    2.24
   11:47     114.48      183585 s     2076.187      77.95    2.73
   17:45     128.25      185908 r     1395.801      86.23    5.19
   09:34      45.97       67158 r      935.067      85.56    0.61
   09:23     113.50      183270 s     2600.852      77.41    9.64
   12:24      51.65       74916 r      804.656      86.18    9.28
   13:30     117.92      183352 s     1809.320      80.39    9.08
   19:23      56.42       89640 s      616.586      78.67    6.77
   11:49      11.29       16205 r      182.767      87.08   10.17
   09:05      34.35       50796 s      744.325      84.53    8.47
   21:12      56.17       76423 r      480.512      91.88    7.53
   01:00      15.33       23930 r     3156.628      80.08    3.11
   03:04      54.60       89731 s     3892.062      76.07    3.81
      06       2.62        4106 r     5071.484      79.88    3.77
   05:15      54.79       89731 s     2277.559      76.32    3.68
      03       2.02        3161 s     7778.530      79.92    2.17
      33       9.42       14680 r     3472.810      80.19    2.31
      00       0.22         219 s     2646.526     127.28    1.81
   19:57     295.16      473489 s     3162.399      77.92    1.85
      10       6.62       10025 r     7841.987      82.54    2.75
      01       0.23         221 s     1092.032     128.96    2.74
      16       6.45       10032 r     1888.591      80.36    2.79
      04       2.06        3243 s     6020.887      79.52    2.62
   01:28      31.29       48532 r     4382.419      80.60    2.62
   07:17     196.34      316072 s     5777.687      77.65    3.86
   01:46      30.14       45786 r     3434.229      82.29    3.26
   01:30      24.73       38405 r     3399.274      80.50    1.80
   02:10      23.46       35633 r     2190.508      82.31    2.61
      44      28.80       46897 s     8441.544      76.76    3.26
   04:51     192.20      316318 s     8671.027      75.95    3.10
      40      11.51       18511 s     3633.437      77.70    2.98
      12       7.17       10963 r     6894.427      81.76    3.04
      12      11.30       18511 s    11418.614      76.32    3.14
      14       7.12       11122 r     6298.740      80.03    3.24
      02       0.92        1412 s     5120.580      81.53    3.41
      14       7.23       11122 r     6184.042      81.24    3.20


        [This RFC was put into machine readable form for entry]
    [into the online RFC archives by Helene Morin, Viagenie, 12/99]





Hicks & Wessler                                                 [Page 6]