On 18/06/2013 7:21 p.m., Nguyen Duc Thien wrote:
Hi folks, I'm requested to build a HTTP proxy server for a pool of 600k subscriber. I intend to use Squid without caching. In the existing system, the average thru-put is something around 450 Mpbs. I have no tool to find out what is the number of concurrent access from the subscribers.
Then you have no tool to accurate gauge the requirement. Nobody here can tell you what the HTTP protocol behaviour of your own subscribers is. The best you will get is a rough educated guess like this one...
Please also read this http://wiki.squid-cache.org/SquidFaq/ConfiguringBrowsers in particular the section about recommended network configuration.
450Mbps -> 9x 50Mbps and one Squid instance can easily handle around 50Mbps on any modern hardware. So you will need 9 or so instances of Squid at one CPU core dedicated to each, YMMV. I would spread this out over several multi-core machines with enough spare cores that one could go down and be worked on while the others take the load (and have a core on the box dedicated to the OS as well). The best way to roll this out without having good pre-knowledge of the HTTP req/sec rates is *slowly*. Once a box appears to be ready to go live push only a small cross-section of your subscribers to it and ensure everything still works well enough. Before increasing that group of subscribers using it over several stages until your whole subscriber base is covered.
Also note: You WILL find that without caching going through the proxy is slower than not going through the proxy.
For this case, what could be the hardware requirement to serve the whole pool of 600k subscribers? Please help. I would prefer IBM platform.
If you have seen the recent thread in "Best OS" it is clear that there is no "best" and any platform you are confident tuning up for the job will do it.
Amos
