taking python enterprise level?...

Roy Smith roy at panix.com
Thu Feb 25 20:21:51 CET 2010


In article <mailman.244.1267114628.4577.python-list at python.org>,
 "D'Arcy J.M. Cain" <darcy at druid.net> wrote:

> The problem had to do with the way TCP/IP works, especially closer to
> the core.  Our provider was collecting data and sending it only after
> filling a buffer or after a timeout.  The timeout was short so it
> wouldn't normally be noticed and in most cases (web pages e.g.) the
> connection is opened, data is pushed and the connection is closed so
> the buffer is flushed immediately.  Our patterns were different so we
> were hitting the timeout on every single transaction and there was no
> way we would have been able to keep up.
> 
> Our first crack at fixing this was to simply add garbage to the packet
> we were sending.  Making the packets an order of magnitude bigger sped
> up the proccessing dramatically.  That wasn't a very clean solution
> though so we looked for a better way.

Interesting, the system I'm working with now has a similar problem.  We've 
got a request/ack protocol over TCP which often sends lots of small packets 
and can have all sorts of performance issues because of this.

In fact, we break completely on Solaris-10 with TCP Fusion enabled.  We've 
gone back and forth with Sun on this (they claim what we're doing is 
broken, we claim TCP Fusion is broken).  In the end, we just tell all of 
our Solaris-10 customers to disable TCP Fusion.



More information about the Python-list mailing list