MTU 1500 gigabit connections translated to 9000 jumbo packets over 10 gigabit connection

Question

I'm testing the frame size by adjusting it from 1500 to 9000 for the link to the storage NAS to optimize the 10 gigabit connection and have seen an increase of speed to almost 300% read speeds.

I would like to further test the stability of this tweak, but what I would like to know is whether there will be a decrease in performance if the clients can only support up to gigabit speed. The setup is the following:

For the test: Dell server with 10 GbE NIC and jumbo packets set to 9000 Direct link to Synology NAS with 10 GbE NIC and MTU set to 9000

In production the following will be applied: D-link 1510 52x gigabit switch with 10 GbE uplink to Synology NAS.

That switch would receive 1500 size packets from the clients (and can't be changed because it also provides the connection to the Internet), and would have the uplink to the NAS set to 9000 jumbo packets, if that's a possible setup.

Will that be a reliable (and possible) setup in the way that gigabit 1500 MTU connections will go to 10 gigabit 9000 MTU to the NAS via the uplink? I know that the other way around that would result in dropped packets, but is it also for this case?

Answer 1

Switching a frame through a different speed port can't change the frame size. The sender uses what it is configured to and what it chooses at that moment. The frame stays that way until it is delivered or dropped (with the expection of tags). (An IPv4 packet may be fragmented while passing a router though.)

Standard frames will go to a port configured for jumbo frames just fine. Jumbo frames delivered to a port with standard frame size will be dropped, increasing the oversized counter.

Best practice is to run jumbos in a separate VLAN to avoid problems. Routing out of that VLAN can put additional strain on the router, so you might want to test thoroughly.

Depending on your hardware, jumbos might not be worth the trouble - the practical throughput increase might be negligible (e.g. TCP could theoretically gain 4.4%). A better reason for jumbos is to remove processing bottlenecks by significantly reducing the frame rate, but modern hardware and offloading features usually make that superfluous.