Quantum| MTU (Jumbo Frames) Configuration Overview in Lattus |
It should never hurt to set the switch to allow a jumbo frames. The hosts can continue to use a 1500 byte maximum transmission units (MTU) if they wish, but if the switch isn’t set to allow jumbo frames, then none of the attached hosts can be configured to use jumbo frames.
In general, if any hosts are configured to use jumbo frames, they should all be configured to use jumbo frames. They should all be configured to the same size of jumbo frames as well.
The reason for using jumbo frames is to reduce the per-packet overload on the host CPU. With Large Segment Offload (LSO), there no longer is any host CPU savings on the sending side, but the receiving side still takes a hit processing the larger number of smaller packets generated by LSO. Large Receive Offload (LRO) can help, but doesn’t fully offset the host CPU processing overhead.
If you run with a mix of hosts configured with and without jumbo frames, things will still mostly work. TCP should detect the mismatch via the TCP MSS option and use the smaller size. Large UDP packets that need to be fragmented at the IP layer may not work when going from the larger MTU host to the smaller MTU host.
The configuration for jumbo frames only matters within the context of a single subnet: i.e., the switch and hosts directly connected to the same switch. You can have jumbo frames enabled on one subnet and disabled on another subnet, and things should work just fine, because the the subnets are separated by routers. (Note: Physically, there can be multiple switches interconnected to provide more ports, but logically it is still one switch, and all ports should be configured to the same MTU size.)
You shouldn’t see IP fragmentation of TCP packets, as TCP will figure out from the TCP MSS options and from Path MTU discovery what is the largest size of IP packet that can be sent that won’t require IP fragmentation. Large UDP packets do use IP fragmentation, and you can’t get around that. If the sending UDP host sends 9000 byte IP fragments and those pass through a router to another network that only supports 1500 byte MTU, the router will re-fragment the packets as they pass through into smaller 1500 byte packets. Then it’s just a question of what is more efficient: having the router do additional fragmentation, or having the source create smaller (but a larger number of) fragments that the router can just directly forward.
This page was generated by the BrainKeeper Enterprise Wiki, © 2018 |