We started running IPv6 at our office about a year ago. I've noticed that services that use IPv6 and run through our router are sometimes very slow. I ran some diagnostics with Wireshark to find that there are large delays between IPv6 packets received, where sometimes frames plus their retransmissions would come through at the same time.
After some troubleshooting, I found two things.
- This condition only occurs when traffic shaping has kicked in and the class matching my service has a queue depth greater than 0.
- If I disable IPv6 on my host and use the same service on IPv4, the packets are not experiencing large delays when congestion occurs.
I have a hypothesis. I wanted to bring it up here in case someone had some ideas that I hadn't thought of yet. Or maybe I've discovered some interesting behavior about how MQC handles dual stack queues.
My thought is that the traffic shaper will service IPv4 packets in a queue first, and then if there are any remaining tokens it will service IPv6 packets. Given the delays I've seen, I believe this may be happening at every traffic shaping Tc interval, thus starving the IPv6 packets in the queue.
I am running a test to split out IPv4 and IPv6 into different queues to see if this makes a difference. To accomplish this, I have created a new policy-map called DUALSTACK and applied it as a child to the service class:
class-map match-any IPV6
match protocol ipv6
class-map match-any IPV4
match protocol ip
policy-map DUALSTACK
class IPV6
bandwidth percent 49
fair-queue 1024
class IPV4
bandwidth percent 49
fair-queue 1024
policy-map OUTPOLICY
class SOMESERVICE
bandwidth percent 20
service-policy DUALSTACK
The router is a 7206vxr G1 running IOS 15.2(4)M5, so HQF is utilized.
