XG-7100 1U Bridging Performance

  • Hi,
    I just purchased an XG-7100 1U to use at a colocation facility with a few servers. I'd like to bridge the 2xSFP+ ports with the LAN switch.

    How much of a performance/throughput hit will there be from bidging 10Gb SFP+ ports if I'm copying data between servers?

    The alternate would be to buy an SFP+ switch, and pay for an extra 1U of space as well as power, which I'd like to avoid if possible.


    EDIT: To be clear, I'm aware that this isn't recommended/ideal. I'm just trying to weigh the performance/throughput cost versus the cost of an extra switch plus 1U of space rental at the colo.

  • LAYER 8 Netgate

    It's all done in CPU, which switch manufacturers avoid like the plague.

    I don't think there are any performance numbers to give you as it is not recommended and has not been tested.

    My advice is to try it and test it and see if it will meet your needs or not.

  • Yeah, I tried it out, and it's quite slow. Getting around 2Gbps as opposed to ~9Gbps on a switch.

    Given this performance, what's the rationale for having SFP+ ports on the appliance? Is it just for media compatibility? If this is the switching throughput, I'd imagine routing throughput will be even lower.

  • LAYER 8 Netgate

    It is a layer 3 router, the layer 2 bridging is more for creating a transparent proxy, not bridging into a "switch." There is no way the CPU is ever going to be as performant as a switch ASIC, etc.

    Also, the links between the SoC and the Marvell switch chips are 2 x 2.5Gb load balanced so I would not expect a single stream to ever exceed 2.5Gb (You got 2Gb) with 5Gb being the absolute overall top end for real-world, multi-stream traffic.

  • Right, I'm aware that CPU switching is much slower than switching fabric. Based on my understanding of the interface links as listed in the documentation, if I'm transferring data from IX0 to IX1 it should never hit the Marvell switch, therefore the 2x2.5Gb limit isn't relevant to my test, right?

    So any throughput bottleneck I'm hitting is just the limit of the CPU, which is where routing is done as well. Given that, I was just wondering why there are 10Gbps SFP+ ports when it'll never get close to being saturated when used as a router.

  • Netgate Administrator

    Yes, you're right it won't hit the switch if you tested between ix0 and ix1. Did you disable filtering on the bridge members?

    As said I don't think we have ever measured the performance of the bridge driver, it's not really intended to be used for that.


  • Yes, I did disable filtering (net.link.bridge.pfil_member=0, net.link.bridge.pfil_bridge=1)

    Am I correct in assuming that switching performance should be better than routing? i.e. if it's only used for layer 3 routing as intended, than throughput would be even lower?

  • Netgate Administrator

    Not necessarily. Most of the same things have to happen in terms of moving data between the two NICs via the CPU. However I doubt the same effort has gone into bridge performance that has routing.


Log in to reply