On 11/29/06, Xuekun Hu <xuekun.hu@xxxxxxxxx> wrote:
hi, I'm trying to predict how many memory should be used with Apache2.2.3 (worker) + Linux2.6.x. This setup is only serving static content, and trying to serve 10000 simultaneous users (about 10000 connections,500 simultaneous requests at the same time). I have several questions below. 1. How many memory is used for holding one connection across software layers (including kernel TCP/IP stack and apache)? 2. How many memory is used for holding one request across software layers (including kernel TCP/IP stack and apache)?
The best way to answer these questions is by experimentation, because they will vary so much depending on your individual setup. Measure total memory usage (using top or equivalent) without apache running. Then measure it again under simulated load and take the difference. I doubt you'll find any difference between the memory needed for a connection and the memory needed for a request. Apache holds one thread for each connection, so there is no memory benefit to idle connections (unless you are using the event mpm).
3. If memory is exhausted and total connections are not exceeded the MaxClients limit (that means Apache will return no mem), what happens if a new connection coming? Apache send "503" or "500" error? And also what happens if a new request coming in a keepalived connection? Apache send "503" or "500"? Does Apache have a pending handling mechanism (I mean holding the connection or request till memory is available, and not sending error message to clients) when out of memory?
In reality, you should never let that happen. If apache exhausts physical memory, the kernel will start to use swap space, which will cause disk thrashing that will completely kill your server. You need to restrict MaxClients so apache never exceeds physical memory. Apache does not send 503s in any default config. This is because, in order to send a 503, you need to have a process/thread available to answer the request and send the error message. But if this process/thread is available, you might as well use it to send the requested content rather than the error message. Apache uses the OS listen queue to handle requests above MaxClients. See the ListenBacklog directive. Joshua. --------------------------------------------------------------------- The official User-To-User support forum of the Apache HTTP Server Project. See <URL:http://httpd.apache.org/userslist.html> for more info. To unsubscribe, e-mail: users-unsubscribe@xxxxxxxxxxxxxxxx " from the digest: users-digest-unsubscribe@xxxxxxxxxxxxxxxx For additional commands, e-mail: users-help@xxxxxxxxxxxxxxxx
|