[Twisted-Python] Twisted performance?

[Tommi Virtanen]

On Mon, Oct 07, 2002 at 09:08:08AM -0700, dave at pythonapocrypha.com wrote:
> Requests per second is only mildly interesting to me - our
> application is all about total throughput. It's essentially a custom
> HTTP server for large files (few hundred megabytes), so I'm looking
> for total throughput numbers for a few high-speed connections as
> well as for many slow (modem or slow DSL) connections. For example,
> on a 800 MHz PIII with a SCSI drive we can currently sustain around
> 300 Mbps total across 100 non-rate-limited connections (pulling the
> data off the disk), and 300 Mbps total across about 1000 or so 150
> Kbps connections. On a faster CPU and no disk access we occasionally
> hit a gigabit per second across 50 or so full-speed connections, but
> my target is only about 250 Mbps per machine, assuming the CPU is in
> the < 1 GHz range. One problem with our implementation is that the
> performance curve drops off too quickly as number of concurrent
> connections go up. I'd love to sustain 250 Mbps across 2000 or more
> connections; for now we compensate by buying more CPU (which isn't
> too big of a problem).

	Umm, you want recent 2.4 kernel, sendfile(2) and
	http://www.kegel.com/c10k.html.
	Maybe throw in an Intel e1000 with TCP segment offloading.

	But maybe that's just the kernel hacker in me.

-- 
:(){ :|:&};: