GigE Question - Aix

This is a discussion on GigE Question - Aix ; All, I believe I know the answer but am interested in unbiased opinions :-) so... For GigE adapters on pSeries/System P running AIX, what is the suggested/correct adapter setting (adapter and port) for speed and duplexing?? Should it be left ...

+ Reply to Thread
Results 1 to 6 of 6

Thread: GigE Question

  1. GigE Question

    All,
    I believe I know the answer but am interested in unbiased opinions :-) so...

    For GigE adapters on pSeries/System P running AIX, what is the
    suggested/correct adapter setting (adapter and port) for speed and
    duplexing?? Should it be left at autonegotiate on both sides or
    "forced" to 1000 Full??

    Also, if an HACMP cluster is NOT using hardware address taekover, does
    the "standard" change???

    TIA

    Pete
    p.s. I'm "building my case" for some changes that ***I*** think are
    required for our environment to solve some intermittent (I believe are
    network switch/adapter related) problems :-) Thx...Pete

  2. Re: GigE Question

    Wylie Coyote schrieb:
    > All,
    > I believe I know the answer but am interested in unbiased opinions :-)
    > so...
    >
    > For GigE adapters on pSeries/System P running AIX, what is the
    > suggested/correct adapter setting (adapter and port) for speed and
    > duplexing?? Should it be left at autonegotiate on both sides or
    > "forced" to 1000 Full??
    >
    > Also, if an HACMP cluster is NOT using hardware address taekover, does
    > the "standard" change???
    >
    > TIA
    >
    > Pete
    > p.s. I'm "building my case" for some changes that ***I*** think are
    > required for our environment to solve some intermittent (I believe are
    > network switch/adapter related) problems :-) Thx...Pete


    Hi,

    if you want to run your Gigabit Ethernet adapter at Gigabit speed you do not
    allways have a choice about the settings, since to my knowledge there are
    Gigabit adapters (adapter drivers) where you can specify 1000_Full_Duplex
    explicitly, while others offer Autonegotiation (for the Gigabit/s setting) only.

    We do use *_Full_Duplex wherever possible and we face problems, if "Auto"
    results in _Half_Duplex media speed running (see entstat). Same applies to HACMP.

    Regards,
    Uwe Auer

  3. Re: GigE Question

    On 08/31/2008 11:00 PM, Wylie Coyote wrote:
    > For GigE adapters on pSeries/System P running AIX, what is the
    > suggested/correct adapter setting (adapter and port) for speed and
    > duplexing?? Should it be left at autonegotiate on both sides or
    > "forced" to 1000 Full??


    It should always match both sides, the more foolproof method would be to
    force it 1000/full.

    > p.s. I'm "building my case" for some changes that ***I*** think are
    > required for our environment to solve some intermittent (I believe are
    > network switch/adapter related) problems :-) Thx...Pete


    An often overlooked criteria is that the switch plays a significant role
    in the performance that can be expected. Even if you have 2 clients that
    are set to 1000/full, the switch may still cause slow performance i.e.
    when it fragments packets.

    There is an IBM document that in essence recommends that you use the
    highest MTU size possible supported by the adapter and network and to
    also ensure you have a homogeneous switch environment with the same port
    speed for all connected clients.

    To put your performance in perspective, I would recommend you use
    something like netperf to test the (pure network) throughput between two
    servers.

    hope this helps!

    Niel

  4. Re: GigE Question

    On 09/01/2008 09:52 PM, Niel Lambrechts wrote:
    > It should always match both sides, the more foolproof method would be to
    > force it 1000/full.


    Sorry, I really meant "on adapters that support it".

    cheers
    Niel

  5. Re: GigE Question

    > > For GigE adapters on pSeries/System P running AIX, what is the
    > > suggested/correct adapter setting (adapter and port) for speed and
    > > duplexing?? Should it be left at autonegotiate on both sides or
    > > "forced" to 1000 Full??


    > It should always match both sides, the more foolproof method would
    > be to force it 1000/full.


    The IEEE standard for Gigabit requires autoneg. It is not an optional
    part of the standard. So, if you have two bits of GbE kit and they
    don't autoneg with one another, one or the other of them is
    fundamentally broken.

    > > p.s. I'm "building my case" for some changes that ***I*** think
    > > are required for our environment to solve some intermittent (I
    > > believe are network switch/adapter related) problems :-)


    What makes it seem switch/adapter related? Are there packet losses
    being recorded in the switch or adapter stats/

    > An often overlooked criteria is that the switch plays a significant
    > role in the performance that can be expected. Even if you have 2
    > clients that are set to 1000/full, the switch may still cause slow
    > performance i.e. when it fragments packets.


    I hate to pick nits with a netperf supporter but

    There is no fragmentation in Ethernet (a layer 2 protocol). There is
    fragmentation in IP (a layer three protocol). Ethernet switches,
    being layer-two devices do not fragment. IP routers, being layer
    three devices, can fragment. Even though there are
    egg-laying-wolly-milk-pig devices out there with switch and router
    functionality in the same sheetmetal, (and even a wireless bridge and
    those evil things called NAT in a "home router") called by only one
    name or even compound names it is important to think of the functions
    independent of their colocation within the same sheetmetal.

    If a switch does not support the MTU used by systems at either end,
    the result is (at least temporary) loss of connectivity when those
    large frames are sent.

    > To put your performance in perspective, I would recommend you use
    > something like netperf to test the (pure network) throughput between
    > two servers.


    Agreed

    rick jones
    --
    firebug n, the idiot who tosses a lit cigarette out his car window
    these opinions are mine, all mine; HP might not want them anyway...
    feel free to post, OR email to rick.jones2 in hp.com but NOT BOTH...

  6. Re: GigE Question

    Wylie Coyote wrote:
    > All,
    > I believe I know the answer but am interested in unbiased opinions :-)
    > so...
    >
    > For GigE adapters on pSeries/System P running AIX, what is the
    > suggested/correct adapter setting (adapter and port) for speed and
    > duplexing?? Should it be left at autonegotiate on both sides or
    > "forced" to 1000 Full??
    >
    > Also, if an HACMP cluster is NOT using hardware address taekover, does
    > the "standard" change???
    >
    > TIA
    >
    > Pete
    > p.s. I'm "building my case" for some changes that ***I*** think are
    > required for our environment to solve some intermittent (I believe are
    > network switch/adapter related) problems :-) Thx...Pete


    Didn't know how else to do this w/o making other posters' replies seem
    irrelevant/ignored :-) so I'm replying to my original post.

    While I agree with all the replies, the 2 that most closely align with
    my understanding are those submitted by Uwe and (more so) Rick Jones.
    Additionally, I researched IBM's web sites and all the links I found
    (related to my question) APPEAR to suggest that IBM also recommends
    leaving GigE (running on SystemP/AIX) at the default, which is Auto
    negotiation (which then implies the switch port requires the same - but
    Auto is the default on the switch port, for most switches anyway, if I
    recall correctly).

    I've now more ammunition in my belt to support my recommendation to
    change the GigE adapters/affected ports to Auto/Auto where I'm currently
    "seat warming" :-P (In case you've not guessed yet,) Yeah, I'm in a
    "prove it" debate with one of my colleagues; that debate started over
    "intermittent" (GigE latency/performance) issues. Most notably so, when
    there is no (heavy) load on the system or disk. In fact the same
    systems have 100 Mb/S adapters set to 100 Full (switch port also) and
    THEY work better than the GigE in their current config (dohhh :-O ).
    All the affected GigE adapters/ports have been manually set to (yup, you
    guessed it) 1000 Full Duplex (sighhhhh :-P )...I'll leave it to y'all to
    surmise who set them that way :-)

    Thanks...I guess I'm not as crazy as I was being led to believe :-)

    Pete

+ Reply to Thread