Scaling CSMA/CD to 1000 Mb/s

Experimental Results

Howard M. Frazier, Jr.
Ariel Hendel
Shimon Muller
Sun Microsystems Computer Company
Internet and Networking Products Group
12-March-1996

Outline

o       Introduction
o       Setup
o       Results
o       Conclusions

Introduction

o       Follow up to presentation made to HSSG on January 11th, 1996
o       Concerns raised about performance with scaled parameters
o       Decided to characterize performance via two methods:
o       Experimentation
o       Simulation
o       Performance characterized in terms of:
        o       Packet transmission rate
        o       Network utilization
        o       Collision frequency
        o       excessiveCollisionError frequency

Setup

o       4 Node 100BASE-FX network
o       4 Sun Ultra 1/170E workstations
        o       167 MHz UltraSPARC uniprocessor
        o       stock Solaris 2.5 kernel
        o       Motherboard 100BASE-T interface
o       Cisco (Grand Junction) 100BASE-FX FastHub 100
o       LANCAST and Canary 100BASE-FX transceivers
o       Various lengths of fiber optic cable
o       ttcp traffic generator

Setup
         [diagram of setup]
         [four Sun workstations connected via Lancast and Canary
          100BASE-FX transceivers with fiber optic spools, to
          Grand Junction fiber optic FastHub]

Setup

o       The fiber optic links were the same length for each station
o       The following link lengths were tested:
        o       2 m
        o       100 m
        o       400 m
        o       600 m
        o       800 m
o       The following parameters were adjusted:
        o       Tx minFrameSize
        o       Rx minFrameSize
        o       slotTime
o       Tx minFrameSize and Rx minFrameSize were used to simulate 
        the behavior of the "Extension Proposal" 

Setup

o       The parameters were stepped through the following values:

        o       64 bytes
        o       128 bytes
        o       174 bytes
        o       256 bytes

o       Two test cases were run for each set of parameters

        o       One repeater with 2 m links
        o       One repeater with the longest links allowed by the parameters

Setup

o       The ttcp test program was used to generate UDP frames
o       Four streams of UDP frames were generated simultaneously
o       A -> B, B -> C, C -> D, D -> A
o       Each station presented 100% offered load
o       Each test was run for 3 minutes, statistics were gathered 
        during the last two minutes of the test
o       Statistics were collected at each station individually, and 
        averaged over the course of the test, normalized to a 1 second 
        interval
o       Statistics are presented as aggregate values
        (summed across all four stations)

Results

o       To interpret tables, use the following key:
o       txpps 	= 	aggregate transmit packets per second
o       rxpps 	= 	aggregate receive packets per second 
o       colps 	= 	aggregate collisions per second 
o       oerrps	= 	aggregate excessiveCollisionErrors per second
o       udput	= 	udp utilization, multiply txpps by 1538
o       colut	= 	collision utilization, multiply colps by slot time
o       Shaded columns of tables represent arithmetic extrapolation,
        not real results

Results
         [table and graph of results for 4 m diameter]
         [last column is shaded]


      diameter       4           4          4           4          4
      slottime      64         128        174         256        512
      txpps       7851.25     7757.53    7705.84     7693.82    7650.00
      rxpps       7851.74     7747.95    7730.31     7681.88    7650.00
      colps       4883.98     2686.02    2043.07     1427.18     800.00
      oerrps       127.18       61.72      44.58       30.29      15.00
      udput          0.9648      0.9532     0.9469      0.9454     0.9400
      colut          0.0328      0.0318     0.0284      0.0315     0.0300
      busy           0.97        0.94       0.96        0.97       0.95
      idle           0.03        0.06       0.04        0.03       0.05
                               
                               
Results
         [table and graph of results for max diameter]
         [last two columns are shaded]

      diameter     200         800       1200        1600       2000
      slottime      64         128        174         256        512
      txpps       7805.66     7681.15    7626.95     7560.00    7440.00
      rxpps       7803.63     7691.94    7634.31     7560.00    7440.00
      colps       4835.01     2641.13    1940.92     1420.00     800.00
      oerrps       123.45       61.25      46.08       30.00      15.00
      udput          0.9592      0.9439     0.9372      0.9289     0.9142
      colut          0.0325      0.0313     0.0301      0.0313     0.0340
      busy           0.98        0.96       0.95        0.95       0.95
      idle           0.02        0.04       0.05        0.05       0.05
                                

Conclusions

o       Increasing the slot time and the minimum carrier duration will 
        not significantly degrade throughput for full size packets
o       Correlation between experiments and simulations is very good
o       CSMA/CD at 1000 Mb/s on 200 meter topologies is 
        technically feasible