What is Real-World Throughput of PCI 64-Bit / 66 MHz Cards? - Storage

This is a discussion on What is Real-World Throughput of PCI 64-Bit / 66 MHz Cards? - Storage ; In theory, a PCI card (not PCI-X, and not PCI Express) that runs on a PCI 64-bit bus at 66 MHz should be capable of driving 512 MB per second. I know in the real world that is fantasy. There ...

+ Reply to Thread
Results 1 to 2 of 2

Thread: What is Real-World Throughput of PCI 64-Bit / 66 MHz Cards?

  1. What is Real-World Throughput of PCI 64-Bit / 66 MHz Cards?

    In theory, a PCI card (not PCI-X, and not PCI Express) that runs
    on a PCI 64-bit bus at 66 MHz should be capable of driving 512 MB
    per second. I know in the real world that is fantasy. There
    are usually bottlenecks everywhere, such as in the I/O chipset or
    achitecture of the motherboard, as well as potential bottlenecks
    on the PCI card itself. My question is what is the real world
    throughput that one can never exceed on such a card, on a
    high-end server that is properly configured?

    To give a real-world example: suppose I have a QLogic 2204F
    fibre channel host adapter with four one gigabit ethernet ports.
    If I could really drive 60 MB/sec onto each of those four ports,
    what kind of throughput would the card itself be limited to?

    In terms of servers, which models of Dell or Compaq servers have
    I/O chipsets that enable the best throughput for standard PCI
    cards, and what is the actual limitation of those chipsets? I
    know back when PCI 32 Bit / 33 Mhz was the only game in town, the
    very best servers could only drive about 20 to 30 MB/sec (against
    a theoretical limit of 132 MB/sec for the PCI bus). So I''m
    expecting that the I/O chipsets still impose great limitations.

    If someone knows the actual limitations for typical PCI-X
    chipsets, I would like to know that number as well.

    I understand the value of having multiple independent buses in a
    good server, but I'm trying to focus in on bottlenecks introduced
    specifically by the I/O chipsets and architecture on any one
    slot. I'm also interested in what kinds of bottlenecks are
    introduced by the typical PCI one gigabit host adapter.

    --
    Will
    Internet: westes at earthbroadcast.com



  2. Re: What is Real-World Throughput of PCI 64-Bit / 66 MHz Cards?

    I don't have all my notes on this at hand, so this will be rough
    numbers, but it should help some.

    You don't really need to worry about bottlenecks in modern chipsets,
    the memory controllers and I/O controllers will be fast enough to not
    neck down a single or even a few PCI slots. Although this could be an
    issue with 5 slots all at 100 or 133 MHz depending on the CPU memory
    utilization.

    What you do need to consider is the PCI bus spec. A 64/66 bus bursts at
    512MB/s, the key is "bursts". Once the bus is obtained by the PCI card,
    and the address and command phases are issued and the data phase starts
    then data flows at 512MB/s. But most cards will only hold the bus for
    64 to 256 bytes before they release the bus. And if there are multiple
    requestors on that same bus then it could be as little as 32 bytes
    before they are forced to release the bus.

    Note that we are not talking about bus slots, but the bus itself. Good
    chipsets like the Intel 7520 have 4 fully independent PCI buses, other
    chipsets share all PCI devices on 2 buses.

    So lets say that are multiple devices on the same bus and they all want
    to use it. Now each device gets 32 bytes transferred (4 data cycles)
    but it took them 3 cycles to set up that transfer and releae the bus,
    so the overall transfer rate that the one card can achieve is almost
    half of the burst rate. And there are chipsets I have measured in the
    past that are actually less than half of burst do to a poor bus grant
    algorithm.

    There are many things to consider in overall system performance when
    you are looking into I/O throughput, and if you need a system optimized
    for I/O then look very closely at the chipset you chose. Not to get too
    involved in the Intel/AMD battle, but AMD is currently very poor in the
    I/O subsystem handling.

    -Jim


    Will wrote:
    > In theory, a PCI card (not PCI-X, and not PCI Express) that runs
    > on a PCI 64-bit bus at 66 MHz should be capable of driving 512 MB
    > per second. I know in the real world that is fantasy. There
    > are usually bottlenecks everywhere, such as in the I/O chipset or
    > achitecture of the motherboard, as well as potential bottlenecks
    > on the PCI card itself. My question is what is the real world
    > throughput that one can never exceed on such a card, on a
    > high-end server that is properly configured?
    >
    > To give a real-world example: suppose I have a QLogic 2204F
    > fibre channel host adapter with four one gigabit ethernet ports.
    > If I could really drive 60 MB/sec onto each of those four ports,
    > what kind of throughput would the card itself be limited to?
    >
    > In terms of servers, which models of Dell or Compaq servers have
    > I/O chipsets that enable the best throughput for standard PCI
    > cards, and what is the actual limitation of those chipsets? I
    > know back when PCI 32 Bit / 33 Mhz was the only game in town, the
    > very best servers could only drive about 20 to 30 MB/sec (against
    > a theoretical limit of 132 MB/sec for the PCI bus). So I''m
    > expecting that the I/O chipsets still impose great limitations.
    >
    > If someone knows the actual limitations for typical PCI-X
    > chipsets, I would like to know that number as well.
    >
    > I understand the value of having multiple independent buses in a
    > good server, but I'm trying to focus in on bottlenecks introduced
    > specifically by the I/O chipsets and architecture on any one
    > slot. I'm also interested in what kinds of bottlenecks are
    > introduced by the typical PCI one gigabit host adapter.
    >
    > --
    > Will
    > Internet: westes at earthbroadcast.com



+ Reply to Thread