IBM TotalStorage DS6000 Series: Performance Monitoring and Tuning

280 IBM TotalStorage DS6000 Series: Performance Monitoring and Tuning

performance in LAN environments, but you may still want to experiment with these

parameters.

The default setting of 64 KB for most Linux configurations is fine for most LANs, but too low for

Internet connections. Set this to a value between 256 KB for T1 lines or lower, and 2 to 4 MB

for T3, OC-3, or even faster connections.

To determine the optimal buffer size for your environment, you can use the following formula:

buffer size =2 *bandwidth *delay

Where bandwidth is the bandwidth of the slowest connection between the server and the

client.

Changing the source and recompiling the kernel for TCP

Go to include/linux/skbuff.h header file in your Linux source directory and edit the values for

the maximum send and receive windows.

#define SK_WMEM_MAX 262140

#define SK_RMEM_MAX 262140

Changing values on a running system

Change the maximum parameters to an appropriate value, depending on your connection

speed.

For a Linux kernel 2.4.x system, add the following lines to /etc/rc.d/rc.local:

echo "4096 65536 4194304">/proc/sys/net/ipv4/tcp_rmem

echo "4096 65536 4194304">/proc/sys/net/ipv4/tcp_wmem

The three values describe the minimum, default, and maximum window sizes used by TCP.

The Linux kernel 2.4.x actually does a good job of adjusting the window size automatically,

depending on network conditions. You simply need to specify appropriate minimum and

maximum values.

8.3 Linux monitoring tools

In this section we discuss the tools available for the supported Linux distributions, that will aid

you in the performance monitoring and tuning activities of your Linux system I/O when using

the DS6000.

8.3.1 uptime

The uptime command can be used to see how long the server has been running, how many

logged on users there are, and gives a quick overview of what average load the server has.

The system

load average is displayed for the last one, five, and fifteen minute intervals. The

load average is not a percentage, but instead the number of processes in queue waiting to be

processed. If processes that request CPU time are blocked (which means the CPU has no

time for processing them), the load average will increase. On the other hand, if each process

gets immediate access for CPU time and there are CPU cycles lost, the load will decrease.

The optimal value of the load would be 1, which means each process gets immediate access

to the CPU and there are no CPU cycles lost. The typical loads can vary from system to

system: For a uniprocessor workstation, 1–2 might be acceptable, whereas you will probably

see values of 8–10 on multiprocessor servers.