Conference 7.286::fddi

242.0. "Performance Problem With DecBridge 500" by CUJO::GAY (James D. Gay - Denver Colorado) Mon Apr 22 1991 15:48

        My customer while doing so performance testing of our FDDI products
        seems to have found a performance problem.  They are seeing a dramatic
        performance hit when the DecBridge 5000 fragments IP packets.
        
        The setup they are using is:

                         +-----------------------------+
                         |            RING             |
                +++++  +++++                         +++++  +++++       +++++
                |   |  |   |                         |   |  |   | Ether | S |
                | B |--| C |                         | C |--| B |-------| U |
                |   |  |   |                         |   |  |   |       | N |
                +++++  +++++                         +++++  +++++       +++++
                         |            RING             |
                         +--------------+--------------+
                                        |
                                      +++++
                                      | D |
                                      | E |
                                      | C |
                                      +++++
                                      
        Where:  B = DecBridge 500, DEFEB-AA, REV A01, Software 2.2, Rom 1.6
                C = DecConcenrator 500, DEFCN-AA, REVA02, Sofwtare 2.3 Rom .0
              DEC = DecStation 5000 with DEC FDDI controller
              SUN = SUN/470
              
        Using theie own testing suite they see the following performance when
        sending packets from the SUN to the DecStation 5000:
        
Record
Size    120     240     512     1020    2040    4080    8192    16380   32760
       +---------------------------------------------------------------------
KBits  +
Per    +6583    8564    10486   9946    9388    6765    8959    8892    7749
Second +

        While the same setup going from the DecStation 5000 to the SUN gives:

Record
Size    120     240     512     1020    2040    4080    8192    16380   32760
       +---------------------------------------------------------------------
KBits  +
Per    +3840    5943    7447    7418    6528    2496    976     860     1936
Second +


       TCP/IP on both machines was modified so that IP could output packets
       up to the largest legal ethernet packet on the SUN and FDDI packet
       on the DecStation 5000.
       
       As can be seen there is a destinct drop in performance going from
       the DecStation 5000 to the SUN when we reach the point where the
       DecBridge 5000 must fragment the packets.  In this case there is
       a 10 times decrease in performance.  Also the DecStation 500 does
       not perform as well before this, but they are not worried about
       this.
       
       Any ideas why this is happening.  I called the CSC but they said
       they could not help me.  They advised me to try this notes file
       for help.
       
       Thanks,
       
       Jim Gay
       (Denver)
       (303) 341-3574

    
    For DEC -> Sun the massive hit doesn't happen when the fragmentation
    starts, because the fragmentation must start happening at the 2040 byte
    point. It seems like the degradation is in proportion to the number of
    fragments. I can see why this would happen, but I don't know why the
    effect should be so pronounced.
    
    The Sun -> DEC figures look a bit odd too - why is the 4080 packet size
    so much worse than its neighbours ?
    
    Trevor

    What does the test comprise of?  I assume no packets are being
    dropped.  Look at the IP fragmentation related counters (around
    the bottom of an ELMS "SHOW LINE 1 COUNTERS" screen) to verify.
    
    If no packets are being dropped by the bridge, the only other thing
    I can think of is that the reassembly process on the SUN is very slow
    when the number of fragments is larger than 2.
    
    You may also want to retry the test with the DECstation 5000 configured
    to output only up to the max Ethernet size.  This will take the bridge
    fragmentation out of the equation.

    Anil

    I'm very concerned that CXO told you to check the notes file. They are
    very aware of the escalation paths available to get you the help you
    need. This note has been forwarded to the US Service Delivery manager
    to see if there is a problem..
    
    In the meantime, I'd at least call CXO back and pressure them a little.
    
    Bryan

There are various possible answers, but you should provide more detail about
exactly what the tests do.  The various explanations I can think of would
explain performance drops, but not at the precise message sizes you mentioned.

	paul