1

I have a group of three Dell N5224SF-ON switches at one of my sites. One port in particular is being used to transfer several long-lived, high-volume streams onto a 1Gb/s port. This traffic should expand to fill the available bandwidth, but instead ticks along around 60Mb/s.

The "input" side of the stream is a 100Gb/s port, and the source flow comes from another switch across a 2x10Gb/s LAG.

We've noticed that the "output discards" value for this port is unusually high (ie non-zero). My research suggests that this means the switch is throwing away a packet because it can't be transmitted as fast as the packets are arriving. My guess is that this is causing the computers involved in the traffic flows to retransmit packets occasionally, causing long (for ethernet) pauses in the flows.

The switch isn't logging anything interesting at all at the moment.

I can't find anything in the OS-10 documentation (I'm running 10.5.3.6) related to non-QOS output port buffers, and because all of this traffic is the same class there's no point in QOS at the moment.

I saw a hint on the internet somewhere that if the switch has a monitor port defined it can cause similar problems to what I'm seeing. This switch does have a monitor port setup, it's pulling in all the traffic across all vlans from the aforementioned 2x10Gb/s LAG and feeding it into a Darktrace appliance through a 10G port. This monitor uplink has output discards that are several orders of maginitude higher than the discard rate on my 1G port.

Is there any reasonable way to diagnose why the output discards are happening? Is the monitor likely to be contributing -- or be -- my problem?

Thank you for any insight you can offer.

3
  • 1
    Switches have tiny buffers.
    – Ron Maupin
    Apr 2 at 20:16
  • What's the bandwidth of the original source? TCP flow? Trying to funnel bursts of 10G+ into a 1G port is bound to cause significant frame loss - even if the whole 32 MB buffer is used, the original flow exhausts the buffer inside 28 ms. With the monitoring going on, the exhaustion is more likely to happen inside a very few ms.
    – Zac67
    Apr 2 at 20:45
  • The flow passes from source through 10G media to get to the 1G bottleneck. However the 1G link only passes 60Mb/s through it, so we are skeptical that this alone is the issue. Apr 2 at 21:00

1 Answer 1

1

I'm not familiar with those Dell switches but I would try:

  • (temporarily) shut down the port mirroring - it's quite possible that it puts significant stress on the frame buffer, in turn causing suboptimal handling of 10G-to-1G funneling with high frame loss
  • check whether the buffer can be configured so that the monitoring causes less stress, possibly using QoS allocations (different classes mean different queues)
  • sometimes buffers are allocated by port group of perhaps 8 or 16 (usually poorly documented), so moving the problem port and the monitoring port further apart might change things
  • check whether flow control can be used - due to its head-of-line blocking tendency and varying implementations it's often hard to predict

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.