Index Journeys
RFC 415 UNKNOWN PERFORMANCE · MEASUREMENT

Tenex bandwidth

Measurement of TENEX host bandwidth over ARPANET

On rfc-editor.org ↗
PUBLISHED
November 1972
AUTHORS
H. Murray
STREAM
Legacy
SAMPLE SIZE
1 million bits
MAX REAL BAUD (BUFFERED)
61 kilobits per second
MIN REAL BAUD (UNBUFFERED)
12 kilobits per second
TTY + LP CPU BANDWIDTH
15-20 kilobytes per second
DOI
10.17487/RFC0415

RFC text

Original text · RFC Editor View on rfc-editor.org ↗ 
Network Working Group                                          H. Murray
Request for Comments: 415                    Stanford Research Institute
NIC: 392                                                29 November 1972


                            TENEX BANDWIDTH

   NIC 11584 (RFC 392) BY G. HICKS AND B. WESSLER AT UTAH RECENTLY
   DISCUSSED THE COST OF USING THE NETWORK.

   I WOULD LIKE TO TAKE THESE SAME TYPES OF NUMBERS AND LOOK AT THEM
   FROM ANOTHER POINT OF VIEW.  WITHOUT CALCULATING COSTS, LET US
   CONSIDER ULTIMATE PERFORMANCE OF THE HOSTS.  I WROTE A SIMPLE TENEX
   PROGRAM THAT SENT ITSELF STRINGS OVER THE NET.  IT USED BIN/BOUT AND
   THE BUFFERED SEND MODE.   IT COUNTED THINGS, BUT DID NOT PROCESS THE
   STRING.  WITH SAMPLES OF A MILLION BITS, THE FOLLOWING RESULTS WERE
   OBTAINED.

   BYTES/SIZE      REAL BAUD       CPU BAUD      CPU NEEDED

     50/36          48.8K           291K          16%
    100/36          46.5K           309K          15%
    200/36          52.7K           322K          16%
    400/36          59.3K           267K          22%
    800/36          61.1K           279K          21%

     50/8           22.0K           70K           31%
    100/8           26.7K           76K           35%
    200/8           29.0K           79K           36%
    400/8           30.0K           81K           37%
    800/8           31.9K           78K           40%
   1600/8           38.1K           64K           59%
   3200/8           41.8K           59K           70%

   THE FOLLOWING WERE OBTAINED BY USING SIN/SOUT WITHOUT THE BUFFERED
   MODE.

   BYTES/SIZE      REAL BAUD       CPU BAUD      CPU NEEDED
     50/36          36K             95K           38%
    100/36          44K             95K           45%
    200/36          49K             92K           53%
    400/36          54K             90K           60%

     50/8           12K             22K           57%
    100/8           16K             21K           75%
    200/8           18K             20K           92%
    400/8           20K             20K           99%
   THESE TESTS WERE RUN WITH NO OTHER LOAD ON OUR SYSTEM, AN EARLY
   VERSION OF 1.29.  THE INPUT AND OUTPUT HALVES OF THE TRANSMISSION
   HAVE BEEN AVERAGED TOGETHER.  I.E. THE ABOVE BAUD RATES WERE
   CALCULATED USING 2 MILLION BITS SINCE THE STRINGS WENT OUT AND IN.

OTHER PIECES OF DATA:

   OUR TTY'S AND LPT HAVE ULTIMATE CPU BANDWIDTHS NEAR 15-20 KB.  AT
   2400 BAUD, A TTY TAKES 15% OF THE CPU.  TELNETS NESTED THREE DEEP ARE
   ALMOST COMPUTE LIMITED WHEN JUST TYPING THINGS OUT ON A 2400 BAUD
   TERMINAL.

   AS A ROUGH CALCULATION, THE BANDWIDTH OF THE TIP TAPE TRANSMISSION
   FEATURE - USING STEVE BUTTERFIELD'S (OF BBN) NETMAG PROGRAM - IS
   ABOUT 7 KILOBITS/SEC.  ONE TAPE OF ABOUT 6,000 TENEX PAGES TOOK ABOUT
   4 HOURS TO TRANSMIT.  THIS WAS A FULL TAPE FROM ETAC TO CCA LATE ONE
   NIGHT WITH NOTHING ELSE RUNNING ON OUR SYSTEM.  ABOUT 30% OF THE CPU
   IS NEEDED TO MAINTAIN THIS BANDWIDTH.  AGAIN THIS IS ONLY ABOUT 20KB
   PER CPU SEC.


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

AI ELI5

On this page
PROBLEMIt is unclear how much throughput TENEX hosts can achieve when sending data over ARPANET and how much CPU they need.
MECHANISMA TENEX program sent strings to itself over the network in samples of a million bits, measuring both real baud rate and CPU bandwidth used under different configurations.
THROUGHPUT DATAReal throughput varies from 22 kilobits per second to 61 kilobits per second depending on byte size and mode; CPU usage ranges from 15% to 99% of capacity depending on the transfer configuration.

§1Measuring Network Throughput

This note measures how fast a TENEX host can send data over ARPANET and how much CPU effort it requires. The author wrote a simple program that sent strings to itself, using either buffered or unbuffered I/O modes, and measured the throughput with samples of a million bits.

Someone tested how fast a computer could send data to itself across the network, counting both speed and effort.

§2Buffered Mode Results

Using buffered send mode with different byte sizes, real throughput ranges from 22 to 61 kilobits per second, with CPU usage from 15% to 40%. Larger byte sizes (400, 800 bytes) achieve better throughput (59-61 kilobits per second) using only 21-22% CPU. Smaller sizes (50-100 bytes) reach about 46-48 kilobits per second with 15-16% CPU. A larger configuration with 3200-byte blocks reaches 41.8 kilobits per second using 70% CPU.

Sending in bigger chunks is faster but uses more CPU. Small chunks are slower and less demanding.

§3Unbuffered Mode and System Limits

Without buffered mode, throughput drops significantly: from 36 to 54 kilobits per second with 36-bit sizes, using 38-60% CPU, and from 12 to 20 kilobits per second with 8-bit sizes, using up to 99% CPU. For comparison, local TTYs and line printers achieve about 15-20 kilobytes CPU bandwidth. At 2400 baud, a single TTY uses 15% of the CPU. Three nested Telnet sessions almost saturate the compute capacity when outputting to a 2400-baud terminal. Network tape transmission via the TIP achieves about 7 kilobits per second, using about 30% CPU.

Without buffering, the computer works much harder and goes slower. Local devices and nested network services can nearly use up all the CPU.