Jumbo frames: worth it, or not

A jumbo frame carries around 9000 bytes instead of 1500. Fewer, larger frames means less per-packet overhead and fewer interrupts. The gain is real but narrow, and the cost of getting it wrong is high.

Where they genuinely help

Storage traffic — iSCSI, NFS, replication — where large sequential transfers dominate. Backup networks. Virtualisation and cluster interconnects. Anywhere a small number of hosts move large volumes across a network you fully control.

Expect single-digit to low double-digit percentage improvement in throughput, and a more noticeable reduction in CPU usage on the hosts.

Where they hurt

Anywhere the path is not entirely under your control, and anywhere the traffic is many small transactions. Web, DNS, voice and ordinary user traffic gain nothing, because the packets were never near 1500 bytes to begin with.

The failure mode

This is the part that matters. Jumbo frames must be enabled consistently on every device in the path: both hosts, every switch, every router, and every VLAN interface. Miss one and you get the classic MTU symptom — small transfers work, large ones hang, nothing is logged.

Worse, the fault often appears months later when someone adds a switch with default settings, and nobody connects the new symptom to the old decision.

If you do it

# verify end to end before trusting it
ping -M do -s 8972 storage-host     # 8972 + 28 = 9000
ip link show | grep mtu
show interface | include MTU

Use a dedicated VLAN or physical network for jumbo traffic so the setting is contained. Document it where someone adding a switch will see it. And test with a real large transfer, not just a ping, before declaring it working.

The honest answer

For a storage or backup network, yes. For a general-purpose network, the gain does not justify the fragility.