IBM SG24-4576-00 Server User Manual


 
which represents the maximum server transactions per second or throughput
rate.
After the peak throughput is reached, the curve begins to slope downward. As
the number of users is increased, the
caching engine
of the operating system
begins to break down. The reduction in the disk cache hit rate is caused by the
increasing amount of data that each additional user requests the server to
access. The ratio of requested data to the size of the network operating system
disk cache increases to a point where the network operating system disk
caching is no longer effective. Furthermore, the disk cache is often storing data
that was requested to be written to disk by application users. This write caching,
when employed, causes write-data to occupy disk cache memory space until it is
finally written to the disk surface. Thus, slow disk write operations can
compound the performance degradation of servers that support write caching
because the amount of useable disk cache memory is reduced by write-data
waiting for service by the slower disk system.
The curve continues to break down until it reaches the transaction rate
sustainable by the disk subsystem. In this state, few I/O requests can be
serviced directly from disk cache and most require direct disk access. At this
point, the server has reached saturation.
The initial slope of the curve is dependent upon how quickly transactions can be
processed by the server, which, in turn, depends primarily on how quickly the
LAN adapter is able to transfer data.
The peak of the curve is the maximum server transaction rate that the particular
server configuration is capable of sustaining for a specific transaction type. The
maximum transaction rate is primarily dependent upon performance of the
network adapter and disk subsystem combination. When the graph flattens out,
performance of the disk subsystem has the greatest influence on overall server
performance.
We can gain a better understanding of how to tune server hardware by studying
the curve dynamics as we modify various hardware components. Changing the
LAN adapter or the disk subsystem can alter the height or width of the graph.
For example, in Figure 141 on page 171, adding a faster network adapter will
increase the initial slope of the graph and also provide additional throughput.
Therefore, the total number of transactions that can be processed per second
will be improved. Furthermore, response time seen by any individual application
user will be reduced. However, the reduction in response time is a limited
benefit. As more users are added, the effects of the improved network adapter
will be offset by increased disk I/O and increased write-data, which reduces the
disk cache hit rate and overall server performance.
170 NetWare Integration Guide