[OpenAFS-devel] nojumbo on server doesn't disable jumbograms on clients

Harald Barth haba@pdc.kth.se
Sat, 15 Apr 2006 14:05:52 +0200 (MEST)


> While tracing a one-way performance problem with our AFS server (MTU 1500), I 
> found that on the client side (MTU 8000), every few packets it tried to send a 
> jumbogram, overriding the negotiated MTU of 1500.

* rx can pack more than one rx message into a datagram. 
* rx can make datagrams > MTU. 
Both these "features" have been called jumbogram. I think
the second one is not making things better with todays
computers (fast enough CPUs) and network structure
(fragmentation and MTU discovery often broken).

> Somebody along the way seems 
> to not fragment these packets, so they just get dropped.

That is broken and often there is not much one can do about it.

> +    rxi_nDgramPackets = 1;

I use to set RX_MAX_FRAGS to 1 (rx_globals.h) when I remember it. All
the rpms we use internally are patched like that.

Harald.