Only 2.7 MB/sec over gigabit connection in Windows XP 64 bit ? (Skybuck's Dream PC 2006)

Discussion in 'Asus' started by Skybuck, Apr 13, 2006.

  1. Skybuck

    Skybuck Guest

    Hello,

    I am testing Pentium III 450 mhz overclocked to 600 mhz<-> AMD X2 3800+

    Windows 32 bit <-> Windows 64 bit.

    The Dream PC only achieves 2.7 MB/sec when it tries to send via the
    NVIDIA nForce Networking Controller.

    I haven't tested the other integrated network chip yet.

    This performance seems quite poor since even the PIII 600 mhz is able
    to achieve 5 MB/sec or so to the AMD... with a special pci gigabit
    networking card.

    The question is what is causing this bottleneck ?

    First I thought it could be my own software... I modified it a bit...
    but I no longer think it's my software... ofcourse it could still be a
    problem with my software so maybe I should test with other software as
    well... but suppose I am correct and it's not the problem with my
    software what else could the problem be ?

    I can think of many possibilities:

    1. Slow integrated network chip.
    2. Windows XP 64 bit networking stack bottleneck.
    3. Bad network card settings.
    4. Bad windows settings/limitations/flow control (?).
    5. Maybe wrong cable ? I don't think so it's a cat 0.6 cable.

    The card in the PIII 450 mhz is a 3COM 3C2000-T card.

    Actually the PIII computer is a very crappy computer... the pci card
    could jump out any moment... it doesn't completely fit properly... but
    still... it seems to work ok...

    The problem seems to be with the Dream PC's hardware or software or
    settings.

    Has anybody else noticed poor networking performance with this software
    or hardware ???

    Any idea',suggestions/comments what the problem might be.. or any tips
    are welcome... ?!?!?!

    Bye,
    Skybuck.
     
    Skybuck, Apr 13, 2006
    #1
    1. Advertisements

  2. Skybuck

    Skybuck Guest

    I forgot to mention one important observation:

    If the packet size/payload size is only around 548 bytes the
    performance is poor:
    * only 2.7 MB/sec from AMD x2 3800+/xp 64 bit to PIII 600 mhz/xp 32
    bit.
    * only 2.8 MB/sec from PIII 600 mhz/xp 32 bit to AMD x2 3800+/xp 64
    bit. However the PIII shows it can send at 8.5 MB/sec at this packet
    size. So the AMD computer seems to be the bottleneck. ???

    If the packet size/payload size is large for example arround 50000
    bytes the performance is quite good:

    * 15 MB/sec from AMD x2 3800+/xp 64 bit to PIII 600 mhz/xp 32 bit
    * 20 MB/sec from PIII 600 mhz/xp 32 bit to AMD x2 3800+/xp 64 bit

    Bye,
    Skybuck.
     
    Skybuck, Apr 13, 2006
    #2
    1. Advertisements

  3. Skybuck

    Skybuck Guest

    I also tried changing some network card/chip settings... no effect.

    When sending and receiving at the same time the throughput even halfs
    to 1.35 MB/sec on the AMD/64 bit xp...

    Not good.

    Bye,
    Skybuck.
     
    Skybuck, Apr 13, 2006
    #3

  4. How are the two PCs connected? If you've got a switch, does it
    indicate a 1000Mb/s link? What about the Windows status display for
    the network connection? Also, what's the CPU load on each box when
    you're running this test? If either box is hitting 100% (or getting
    close), that's going to be a limit (and a better NIC and/or driver can
    help there).

    If you're using a direct connection, note that this is not 100%
    reliable for GigE, and that you need to use a straight-through cable
    and *not* a cross over cable like you did for 10/100Mb Ethernet. If
    you use a crossover cable you'll at best negotiate back to 100Mb mode.
     
    robertwessel2, Apr 13, 2006
    #4
  5. Skybuck

    bluethought Guest

    Have you tested entirely using just this pair of computers ? If I were
    you, I'd do some tests with more PCs. What you see as the AMD PC being
    unable to send faster might very well be just the PIII receiving slower
    and buffers getting saturated.

    I have an AMD 3800+ X2 like yours, running win2k3 x64, and I'm pretty
    sure I've hit better speeds than yours...on the internet!
     
    bluethought, Apr 13, 2006
    #5
  6. You expect performance to go down if packet size or window size goes
    down. You encounter different limits in a communications session. Raw
    bit rate of the medium is just one. Another is frame rate limit of the
    hosts. Another is window size, in case ACK frames are used in the
    protocol.

    Search under Nagle Algorithm (RFC 896) for an interesting discussion on
    the latter consideration.

    Bert
     
    Albert Manfredi, Apr 13, 2006
    #6
  7. Skybuck

    Skybuck Guest

    I have isolated the problem I think.

    The following service is necessary for better performance:

    "ForceWare Intelligent Application Manager (IAM)"

    When this service is running the UDP throughput increases too:

    20 MB/sec. (AMD X2 3800+ to PIII 600 mhz ) (payload size 548 bytes)
    6.8 MB/sec (PIII 600 mhz to AMD X2 3800+) (payload size 548 bytes)

    32 MB/sec. (AMD X2 3800+ to PIII 600 mhz ) (payload size 32000 bytes)
    21 MB/sec (PIII 600 mhz to AMD X2 3800+) (payload size 50000 bytes)

    That's much better ! =D

    I have the following services disabled because I don't think they are
    necessary and especially the apache service seems a bit dangerous:

    "ForceWare IP Service"
    "ForceWare user log service"
    "ForceWare Web Interface" (Apache!! <- dangerous !)

    Bye,
    Skybuck.
     
    Skybuck, Apr 13, 2006
    #7
  8. Skybuck

    Skybuck Guest

    Thanks for the response... I think I have solved all or part of the
    problem... see my other posts ;)

    What is the maximum throughput you have seen at lan ?

    Bye,
    Skybuck
     
    Skybuck, Apr 13, 2006
    #8
  9. Skybuck

    Skybuck Guest

    Hello,

    Thank you too for the quick response too.

    I think I have already solved the problem...

    See my other thread about the solution involving a necessary NForceWare
    service for better throughput... kinda strange huh ? ;)

    I will answer your questions just for kicks and some more info ;)

    They are connected via a patch through cable... so not a cross over
    cable ;)

    So there is no switch...

    The status is 1 gigabit connection.

    AMD CPU load is 7 procent when receiving 20 MByte/sec from PIII.
    (Payload size 32000 bytes)
    PIII CPU load is 100 procent when sending 20 MByte/sec to AMD. (Payload
    size 32000 bytes)

    AMD CPU load is 30 procent when sending 33 MByte/sec from PIII.
    (Payload size 32000 bytes)
    PIII CPU load is 100 procent when receiving 33 MByte/sec from AMD.
    (Payload size 32000 bytes)

    One further note/observation is that when the AMD tries to send 40
    MByte/sec and the throughput is only 33 MByte/sec the AMD CPU load is
    only 20%.
    That's kinda funny... so when it's sending too much the CPU load is
    actually lower... kinda strange... but probably some kind of software
    thingy... no biggie.

    So far I think I can be quiet happy with this performance. The PIII
    performance is still a bit low when the packet/payload size is around
    548 bytes... Only 6.8 MB/sec or so.

    Still these speeds are much better than the crappy 2.7 MB/sec...

    Some day I'll also test the speed of both integrated network
    controllers on the AMD... I'll simply loop the cable around and connect
    them both hehe.

    But for now I need the internet connection... so I am not going to do
    that right now.

    Though just out of curiosity... what is the fastest throughput you have
    seen on a LAN/Gigabit connection ? ;)

    Bye,
    Skybuck.
     
    Skybuck, Apr 13, 2006
    #9
  10. Skybuck

    bluethought Guest

    I haven't tested much on gigabit lans...with multiple simultaneous
    connections I've maxed out 100mbits.
     
    bluethought, Apr 14, 2006
    #10
  11. Skybuck

    ohaya Guest

    Hi,

    I'm curious how are you doing your speed testing/measurements?

    I did a lot of benchmark testing on GigE a couple of years ago. This
    was with several different types of server-class machines, including
    some IBM blades. Some of this was over a cross-over cable, and some via
    a GigE switch (in the case of the IBM blades).

    This testing was primarily "pure" network throughput, and one of the
    main tools I used was "iperf":

    http://dast.nlanr.net/Projects/Iperf/

    This takes any disk I/O out of the equation.

    Using iperf, and also with several other tools, I was able to achieve
    nearly (maybe ~80-90%), but not quite, wirespeed.

    Jim
     
    ohaya, Apr 14, 2006
    #11
    1. Advertisements

Ask a Question

Want to reply to this thread or ask your own question?

You'll need to choose a username for the site, which only take a couple of moments (here). After that, you can post your question and our members will help you out.