Subject: Re: kernel ip_randomid() and libc randomid(3) still "broken"
To: Dennis Ferguson <dennis@juniper.net>
From: Jonathan Stone <jonathan@DSG.Stanford.EDU>
List: tech-net
Date: 11/26/2003 14:38:11
[.... rant about Internet growth ]

> I'll bet, however, that 3 years from now you'll
>have forgotten this argument and will be continuing to use the then
>"*not acceptable*" 16-bit ID IPv4, without having a lot of difficulties
>with it in practice.

I think you're understating a key point here: link bandwith goes up in
factors of 10, and does so much less frequently than annual
cumulative-bandwidth figures, or Moore's-law increase in CPU speed.

I, personally, wont think about it much again, until my NFS traffic
starts using 10 Gbit links. At that point, I (personally) will look
very closely into the drop rates on UDP traffic. 3 years from now I
fully expect I wont use UDP/32k read/writge mounts anymore.

(I have some small hope for 16k+epsilon jumbo frames, which
would mitigate the marginal overhead of TCP processing considerably.
Even 9000-byte frames helps)


>> Then consider that 10Gbit Ethernet is already on the market, and set
>> to fall in price once 10GbE-CX4 becomes available.
>
>Exactly.  If a 12k ID space is "*not acceptable*" at 1 Gbps then we're
>already doomed at 10 Gbps even with the full 16-bit space available.

[...]

For NFS-over-UDP-over-1500-byte-Ethernet: Quite possibly, yes.


[...]

>Given that we're operating solely in the realm of opinions,


Not as much as you seem to think. Try visting an NFS server vendor and
ask them about what they did a year or two back when Linux clients ate
*all* their packet buffers with incomplete UDP fragments (some
fragments were dropped by the source, from its transmit queue).  Ask
those same vendors whether they still recommend NFS over UDP.

I've enountered this myself with Linux 100Mbit clients to a gigabit
server.Two misbehaving clients ate all the server's buffers (32 Mbytes
of cluster mbufs) in a vey few tens of seconds. l havent set out to
deliberately reproduce this with gigabit clients, but the prospect of
large numbers of incomplete UDP NFS-write datagrams sitting in server
buffers for tens of seconds (and thus vulerable to reassembly
collisions) is a very real one.