Posted on January 8, 2019 9:46 pm
 |  Asked by Frank Bulk
 |  383 views
0
0
Print Friendly, PDF & Email

We have two Arista 7504’s, each with a downstream-facing LAG into the same Cisco 7600-series router. Both were flat-lining around 10 Gbps, so thinking that perhaps it was the downstream router, we spread the traffic around by adding a fourth 10G to each LAG. Yet we’re flat-lining around the same value — adding that fourth LAG member made no perceptible difference.
There are no deferred or output discards on the port-channel or individual ports, so my remaining thought is that there is some kind of rate-limiting going on in the Aristas that we can’t see.
Is there anything that needs to be done to a port-channel to get beyond the 10 Gbps mark? Are there any other counters we should be looking at?
If the issue was the downstream router I would be expecting to see the Aristas keep pushing traffic out but ingress discards on the downstream router … but I’m not seeing that.

Attachments:
0
Posted by Alexis Dacquay
Answered on August 14, 2020 9:28 pm

Hi Frank,

Firstly, Arista switches are line rate. That is not the bottleneck. A 7504 can drive Tbps.

Secondly, what is the problem? Are your services sending 40G traffic in? Do you have drops or TCP retransmits on the servers? Or maybe your services don't generate/pass more than what you see? Why is flattening at 10G a problem?

What MTU do you have?
At default 1500B, you cannot reach ~95% throughput because of the network header inefficiencies.
At lower average (iMix, etc), then one's network can easily have maximum efficiency of 80-90%. That is running at 100% line rate on a 10G link, and getting 8-9Gbps of traffic. That is normal.
If you applications need to send lot of data, then increasing the MTU to jumbo frame size (9000B and more) means your header-vs-payload efficiency gets near 99.9%, so you can get ~10Gbps off a 10G link.

Now, if your diagram shows anything slightly above 10G, then it means you use more than 1 link.

Please look at your traffic distribution per physical member, not just the aggregated throughput.
show port-channel summary
show interfaces e1-4 counters rates

It will show you exactly how much throughput is passing in/out of each physical interface.

Please consider all these useful counters:
s7057#show interfaces e1-4 counters ?
bins Packet length bin counters
delta Interface counters delta
discards Packet discard counters
errors Error counters
half-duplex Half-duplex error counters
incoming Incoming counters
ingress Ingress counter information
ip IPv4, IPv6 counters
outgoing Outgoing counters
queue Queue counters
rates Input/Output rate counters

Maybe there is no problem, or is there?

Regards,
Alexis

Post your Answer

You must be logged in to post an answer.