Re: IceBreak as an alternative to Apache?

On 12/31/2010 2:38 PM, Aaron Bartell wrote:

I am not sure what the max clients is, but I do know that when I have done
testing with JMeter I hit the TCP layer bottleneck before I hit the OS job
limitation. Actually, that was a test for stateless transactions. I
basically spun up a couple hundred Apache jobs and hit it hard with JMeter.
That's when things started falling apart and I had to call IBM. They of
course told me it was my software, so I had to put together a bare bones CGI
app and prove it wasn't. Then they got one of the higher level devs on and
they pretty much determined it was at the TCP layer. You can adjust the TCP
layer settings, but they recommended against it.

The number I've seen in the literature is 256. That's the default number and you need to do some finagling to increase that number. This seems to correspond to what you're saying - a couple of hundred jobs and things go casters up. What happens, of course, is that requests get "queued" but it depends on the browser side whether or not that queuing gets handled gracefully. Too long a wait and the browser's going to choke.

I know a million connections is a lot, but I am dreaming big for some social
media type applications I have up my sleeve :-) I guess if I ever did get
over a couple thousand concurrent connections that I would want to have
multiple servers at different geographic locations because it would probably
mean I hit the jackpot! ;-)

Exactly. At that point you start to consider scaling horizontally. You might want to have a front end that handles the requests and calls the back end. This is also a good idea to deal with DDOS attacks (just in case you get on the bad side of Wikileaks <grin>). Seriously, though, if you can design your application so that you have a very thin UI layer you can scale out to lots of UI servers which all feed back to one business logic server. It's a pretty solid architecture.

Joe