Before deploying any application to production servers it’s always good to check if the application won’t harm the environment. It’s good to check at least CPU usage & I/O load (local & network). That’s where troubles come from in large scale environments.

The CPU usage for one process can be easliy measured using sar command (SYSSTAT tool) with -x option:

sar -x pid.

Choosing one second monitoring interval will generate lots of results but only this way we can capture any CPU peaks generatedy by a process.

Network I/O Load can also be easily monitored using for example tcpdump.

The biggest problem is with checking the I/O Load of local disks. Linux system counts internally I/O load of a whole system, but not of separate processes. I tried to get an average I/O load generated by the application.
Procedure for one host:

1. Start the application.
2. Run iostat for defined amount of time. Get results: results0
3. Stop the application
4. Run iostat for defined amount of time. Get results: results1
5. Compute change in I/O load for a host: results=results0-results1

The problem with this approach is that everything can affect the I/O Load during measurements time. I tried to use statistcs to solve it. I assumed that when I run this procedure on larger number of hosts and ranodimze procedure startup time, I will be able to extract I/O Load which was generated by the tested application.

I ran this procedure on random time on each of 450 hosts 2 times. Here are results:

It looks like the application generates 0.8 increase in tps on average.

Bookmark to: