[Top] [All Lists]

Re: productivity?

2011-08-22 17:02:43

Steve Atkins wrote:
Measuring actual costs (what is the resource cost of session setup and 
teardown, and of maintaining an open session, for well-implemented, mainstream 
MTAs?) would seem to be something to do before discussing such a document in 
much detail.

I've been measuring this for my own product (an SMTP proxy) mostly in the course of designing more rigorous test tools. (An unbreakable toy is useful for breaking other toys....)

Not surprisingly, at 1 second between messages, the load on my SMTP receiver is dramatically lower (and throughput markedly higher) when senders hold the connection open; doubly so for TLS, where the certificate exchange dominates. Equally unsurprising, when the sender holds the connection for 5 minutes and more, I'm having problems with hitting my limit on concurrent connections. I'm still looking for where the "sweet spot" is between. At present, I blindly 421 any sender that's been idle for 3 minutes, but that's little more than a SWAG. I'd really like to implement throttling based on a sender's actual load, not just on sockets held open; but I don't have the data to cost justify that. Yet.

One issue not raised here: ephemeral port exhaustion. It's a rather specialized problem, more of an issue for Windows folks than Linux, and mostly an issue for routers, proxies, and similar devices that tend to funnel a large volume of mail to smaller number of destinations. But it's a real problem, one where reusing existing client connections can provide a complete and permanent resolution.


<Prev in Thread] Current Thread [Next in Thread>