tech-net archive

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index][Old Index]

Re: problems in TCP RTO calculation



On Sat, Mar 12, 2011 at 12:37:00PM -0500, Thor Lancelot Simon wrote:
> 
> Typical LAN latencies today are on the order of 1/20 the smallest value
> we can represent (16.5ms).

What about atypical LANs - eg two systems connected by MII crossover?
Or MII (SERDES?) crossover via a small GE switch?
ie when the host MAC and target MAC are all on the same PCB.

We (at work) nearly did our own GE switch in an fpga (we found a
commercial part (from marvell IIRC) that didn't need PHYs in the end
and used that to save the development effort.
There wasn't the power budget available to use back to back GE PHYs!

The elapsed time for packet is probably only slightly greater than
twice the transmission time - not long for small packets.

We also have atypical traffic patterns - I've had problems with
an external switch failing to process 30000 packets/sec on one TCP/IP
connection! (with almost no other traffic).
This traffic isn't really synchronised between the send and receive
sides (it is data packets from multiple channels of a slower network)
and I've seen problems (linux both ends) with the 0 RTT making it
always be doing 'slow start'!

NFI how NetBSD would fair here, but a lot of the code seems to be optimised
for FTP (or similar bulk transfer) while giving something half reasonable
for command-response and terminal work.
Any other traffic pattern tends to cause grief!

        David

-- 
David Laight: david%l8s.co.uk@localhost


Home | Main Index | Thread Index | Old Index