Re: switch vs. hub for hobbyist cluster - VMS

This is a discussion on Re: switch vs. hub for hobbyist cluster - VMS ; In article , helbig@astro.multiCLOTHESvax.de (Phillip Helbig---remove CLOTHES to reply) writes: > In my hobbyist cluster I have 2 VAXes and 1 ALPHA as full-time members. > A big ALPHAserver 1200 boots as a satellite when I need the computing > ...

+ Reply to Thread
Results 1 to 7 of 7

Thread: Re: switch vs. hub for hobbyist cluster

  1. Re: switch vs. hub for hobbyist cluster

    In article , helbig@astro.multiCLOTHESvax.de (Phillip Helbig---remove CLOTHES to reply) writes:
    > In my hobbyist cluster I have 2 VAXes and 1 ALPHA as full-time members.
    > A big ALPHAserver 1200 boots as a satellite when I need the computing
    > power to run CSWB. The VAXes are definitely 10 Mb/s, the 1200 can do
    > 100 I believe and I am not sure about the DEC 3000/600 (the older
    > ALPHA). Occasionally other VMS machines are connected up for test
    > purposes, and on the same LAN there is a PC which boots quickly which I
    > occasionally use (dons flame-proof suit) when I need really quick
    > internet access with a modern browser (as the 1200 takes a while to
    > boot).


    I'm failry sure you 3000 600 has a 10Mb/s NIC. So only you're PC and
    1200 could take advantage of the 100Mb/s switch.

    Switching over dumb hub helps, but you cluster traffic and some of
    the other protocols you're using is broadcast, so you won't see
    much improvement going from 10Mb/s hub toi switch.

    I have a 10/100 switch because it was cheap, it was my first, and my
    Macs could all do 100. Now my DS10L does 100. But broadcast makes
    all the lights blink.


  2. Re: switch vs. hub for hobbyist cluster

    You're not going to see magical improvements to your network by going to
    a swicth. But logically, you will have improvements.

    For one thing, if you have 2 machines capable of 100, then they will
    finally eb able to talk to each other at normal speed and not let the
    slower vax slow them down.

    On the other hand, a switch is a "store and forward" device which
    introduces some latency, especially at 10mbps. (swicth receives packet
    fully, then moves it to each of the designated destination port queues
    where the packets are sent out. So to send 1500 bytes, it will take
    twice the transmission time (time for switch to receive packet, and time
    to send it).

    On a hub, the signals are being duplicvated, so the ports receive the
    bits at the same time as they are being sent.

    One big improovement a switch provides is ability to have some nodes at
    full duplex even those older nodes are not.

    So your alpha1 could be receiving data from alpha2 while it is sending
    data to vax1. And it doesn't have to worry about vax1 being busy because
    the swicth wuill accept the packet and deliver it to vax1 whenever vax1
    is ready to listen (since vax are half duplex).

  3. Re: switch vs. hub for hobbyist cluster

    On Aug 29, 9:48 pm, JF Mezei wrote:
    > You're not going to see magical improvements to your network by going to
    > a swicth. But logically, you will have improvements.
    >
    > For one thing, if you have 2 machines capable of 100, then they will
    > finally eb able to talk to each other at normal speed and not let the
    > slower vax slow them down.
    >
    > On the other hand, a switch is a "store and forward" device which
    > introduces some latency, especially at 10mbps. (swicth receives packet
    > fully, then moves it to each of the designated destination port queues
    > where the packets are sent out. So to send 1500 bytes, it will take
    > twice the transmission time (time for switch to receive packet, and time
    > to send it).
    >
    > On a hub, the signals are being duplicvated, so the ports receive the
    > bits at the same time as they are being sent.
    >
    > One big improovement a switch provides is ability to have some nodes at
    > full duplex even those older nodes are not.
    >
    > So your alpha1 could be receiving data from alpha2 while it is sending
    > data to vax1. And it doesn't have to worry about vax1 being busy because
    > the swicth wuill accept the packet and deliver it to vax1 whenever vax1
    > is ready to listen (since vax are half duplex).


    JF, are you aware of "cut through" switching? Once you have the packet
    header for an incoming packet you can start the work. You identify the
    outgoing port, and if the speeds match you can start forwarding the
    data. You risk forwarding packets with bad checksums, but in return
    you improve the latency a bit. The opposite of "cut through" is "store
    and forward", which means the whole incoming packet is processed
    (including checksum) before being forwarded (or, if the checksum is
    wrong, it'll be ignored).

    Some switches can only do one mode, some do both. Even some dirt cheap
    stuff claims to be able to adaptively switch between the two depending
    on traffic (eg http://investor.netgear.com/ReleaseD...leaseID=118752
    - note the hilarious prices at the end, an FS508 introduced at $1600
    is these days probably under $50, if available - I paid 30 or so for
    mine a couple of years back).

    Something else to think about is whether the ability to set up a port
    to see "all" the network traffic is required; sometimes it's helpful
    for techies to be able to do that kind of thing, most folks wouldn't
    care. Typical SoHo switches don't do that, whereas it's an inherent
    part of basic "hub" (I hate that word, it's too ambiguous)
    functionality.

    That being said, for Philip's stated current requirements, his
    existing hub would seem to be more than adequate for *most* aspects.
    The main machines are all 10Mb max, and the chances of them being able
    to see a performance benefit from full duplex 100Mb switching seem
    small, as the NICs are constrained to 10Mb and I'd *assume* (!)
    there's not much benefit in two parallel sets of 10Mb connections.
    Only the AS1200 and the PC can get the full 16Mb of the DSL
    connection, and they aren't the core machines.

    I don't remember any mention of cables in Philip's post. Current
    switches are likely to be twisted pair, so if the existing kit doesn't
    have them, converters will be required and may not be cheap. If no
    adapters are required, or they already exist, then for the cost of a
    current equivalent of something like an FS508, it might be worth just
    buying one and trying it. If nothing else it could be used as a spare
    for the hub (unless one already exists).

  4. Re: switch vs. hub for hobbyist cluster

    In article <48b860d5$0$4543$c3e8da3@news.astraweb.com>, JF Mezei
    writes:

    > On the other hand, a switch is a "store and forward" device which
    > introduces some latency, especially at 10mbps. (swicth receives packet
    > fully, then moves it to each of the designated destination port queues
    > where the packets are sent out. So to send 1500 bytes, it will take
    > twice the transmission time (time for switch to receive packet, and time
    > to send it).


    It's probably at the per-cent or per-mil level, but perhaps it would be
    best to put the VAXes on a hub and connect that hub to a switch
    containing the rest of the connections.


  5. Re: switch vs. hub for hobbyist cluster

    Phillip Helbig---remove CLOTHES to reply wrote:

    > It's probably at the per-cent or per-mil level, but perhaps it would be
    > best to put the VAXes on a hub and connect that hub to a switch
    > containing the rest of the connections.



    Nop. Because of the half duplex nature, you are better off with
    everything on a switch. This allows the 2 vaxes to talk at the same time
    to some 3rd party or even to themselves.

    Also, remember that the switch will not bother VAX2 with packets
    destined to VAX1, so VAX2 has far fewer collisions that delay packets.

  6. Re: switch vs. hub for hobbyist cluster

    Bob Koehler wrote:
    > Switching over dumb hub helps, but your cluster traffic and some of
    > the other protocols you're using is broadcast, so you won't see
    > much improvement going from 10Mb/s hub to a switch.


    OpenVMS Clusters don't use broadcast. They do use Multicast to a
    specific multicast address (based on the Cluster Group Number) for Hello
    packets (which are sent every 1.5 to 3 seconds from each LAN port
    enabled for cluster traffic, and are used for path discovery and for
    tracking path status and latency and packet size capability) but all the
    real inter-node cluster traffic is unicast.

    LAT also does not use broadcast, but uses Multicast to a specific
    multicast address about once a minute for service advertistments. All
    the real data is carried via unicast packets.

    As I understand it, the TCP/IP protocol may use some Ethernet broadcasts.

    I would advise trying an inexpensive switch. The cost is so low these
    days that it's worth trying. Here I've seen 5-port 10/100 switches for
    US$10 or even less.

    You can use the LOCKTIME.COM procedure from the V6 OpenVMS Freeware CD
    to measure lock request latencies before and after the introduction of a
    switch, and get an idea of the difference that way.

    Shadow copies would certainly be helped by going from 10 megabits to 100
    megabits. With Gigabit Ethernet you can also use Jumbo Frames to do
    shadow copy work with fewer packets, and thus lower CPU interrupt-state
    load on the hosts.

  7. Re: switch vs. hub for hobbyist cluster

    In article , Keith Parris
    writes:

    > I would advise trying an inexpensive switch. The cost is so low these
    > days that it's worth trying. Here I've seen 5-port 10/100 switches for
    > US$10 or even less.


    I think I'll spend about EUR 80 on a 16-port 10/100 switch. All of the
    8 ports on my hub are full now (3 active machines, uplink to the DSL
    connection, laptop, 1200 satellite, port for test machines, and
    currently inactive 3000/300 LX) so that should give me some extra room.

    (By the way, although I never INSTALLED VMS 7.3-2 on the 3000/300 LX
    (but rather on the 3000/600 then swapped in the system disk), I actually
    RAN 7.3-2 on it, even though it has only 48 MB. However, when that
    machine had the TCPIP cluster alias, the resources were maxed out. I
    really like its small size, though.)


+ Reply to Thread