Repost: rt-ac66u jumbo frames size?

[hizzo3]

So I've been setting up XenServer for the entire weekend and trying to figure out why my iSCSI wasn't reading right when I bumped the MTU to 7K. So if I use my GS108T, the jumbo frames work without issue. The moment I pass traffic through the Asus router, it fails (Packet needs to be fragmented but DF set when pinging >4000b frames). My iSCSI box 4k frames are 4192, which fail when pinging.

Does anyone know what size jumbo frames the router supports? The iSCSI unit I am using is running 4k sectors, so I am hoping I can push the 7K through so I can push an entire sector+some through per packet, but doesn't seem to be so. the jumbo frame support is flipped on. I've even tried going in via SSH and changing the eth1/2 to 9k packets to see if that fixes it. I've googled it till I'm blue in the face, but all I see is that jumbo frame support - Check! vs any real data on what sizes. I'm running merlinwrt on the box, up to date when checked 2 weeks ago (bonding/LAG will be used later). Hopefully someone can clue me in.
Thanks.

 

[poisonsnak]

I was just playing around with this the other day. My RT-AC66U will pass 9K frames on its Ethernet switch, but not over WiFi, or the WAN. I was pretty sure this would be the case, my WiFi client (Intel 7260AC) doesn't have a Jumbo Frames setting in the driver (does WiFi even support Jumbo Frames...?) and I don't think many if any ISPs support Jumbo Frames (I don't know if the WAN limitation was my ISP or not - I just tried to ping my ISP's DNS server).

I'm running stock ASUS 3.0.0.4.376_3602 and Jumbo Frames is turned on under LAN. Both PCs were running Realtek RTL8111 NICs and set to 9K frame size in the driver. Hope this helps.

As a side note, I didn't really notice any performance increase using iperf between the two machines, so I put it back. This is between two new-ish PCs though so if your iSCSI box has limited CPU power I could see it being a benefit.

 

[jegesq]

I've played around with it but never kept it on. Jumbo frames can slightly raise efficiency of Ethernet by reducing overhead (larger payload, fewer CPU cycles), as for example with TCP over IPv4. But one problem is that many ISP's don't even support Jumbo frames, and even if they do, using them can adversely impact latency, especially with lower bandwidth connections.

Does anyone know whether Jumbo frames have any impact (positive or negative) when using IPv6?

 

[hizzo3]

Thanks. I'll have to go back and recheck everything.

My iSCSI box is running a Marvell 700 MHz processor, so raid 5 across 4 drives + TCP/FCS does drag on the processor. I see around a 10 MByte/s difference on my initiator box over a 1Gbit connection.

For IPv6, I can additional benefit because of the larger frames, but I'd have to go back and check my documentation by how much. I have ipv6 disabled across my network, but plan to implement it at a later date.