Siege - running a stress test benchmark
- by morgoth84
I need to do a benchmark test of a HTTPS server using Siege, to see how it behaves under massive load. I'm initiating tests from another machine which is quite powerful and it is connected to the same physical switch the server is connected on.
But when I initiate a test, I can't get it to make more than 170 requests per second. With this load the server's CPU usage is at 15-20% and the average response time for a request is approx. 0.03 seconds. Load of the client machine is approx. at 10%.
So, I gradually increase the number of users in Siege (the number of worker threads) and request rate linearly increases up to 170 reqs/sec, but it never gets over it. No matter how many more worker threads I start, the load on the server is never more than 20% (and the client's load also doesn't increase any more).
How can I overcome this?
I've googled a bit and found out that after a request is completed, a socket associated with one ephermal port remains in WAIT_TIME state for some time during which it can't be reused. I tried to overcome this by doing these things:
sysctl -w net.ipv4.ip_local_port_range="1024 65535"
echo 1 > /proc/sys/net/ipv4/tcp_tw_recycle
Oh, and the client machine is a Linux (RedHat, I think, but I'm not sure). Any help would be appreciated.
