Performance tuning is primarily a matter of resource management and correct system-parameter setting. Tuning the workload and the system for efficient resource use consists of the following steps:
There are appropriate tools for each phase of system performance management (see Appendix A. Monitoring and Tuning Commands and Subroutines). Some of the tools are available from IBM; others are the products of third parties. The following figure illustrates the phases of performance management in a simple LAN environment.
Figure 1-2. Performance Phases. The figure uses five weighted circles to illustrate the steps of performance tuning a system; plan, install, monitor, tune, and expand. Each circle represents the system in various states of performance; idle, unbalanced, balanced, and overloaded. Essentially, you expand a system that is overloaded, tune a system until it is balanced, monitor an unbalanced system and install for more resources when an expansion is necessary.
It is essential that all of the work performed by the system be identified. Especially in LAN-connected systems, a complex set of cross-mounted file systems can easily develop with only informal agreement among the users of the systems. These file systems must be identified and taken into account as part of any tuning activity.
With multiuser workloads, the analyst must quantify both the typical and peak request rates. It is also important to be realistic about the proportion of the time that a user is actually interacting with the terminal.
An important element of this identification stage is determining whether the measurement and tuning activity has to be done on the production system or can be accomplished on another system (or off-shift) with a simulated version of the actual workload. The analyst must weigh the greater authenticity of results from a production environment against the flexibility of the nonproduction environment, where the analyst can perform experiments that risk performance degradation or worse.
Although you can set objectives in terms of measurable quantities, the actual desired result is often subjective, such as satisfactory response time. Further, the analyst must resist the temptation to tune what is measurable rather than what is important. If no system-provided measurement corresponds to the desired improvement, that measurement must be devised.
The most valuable aspect of quantifying the objectives is not selecting numbers to be achieved, but making a public decision about the relative importance of (usually) multiple objectives. Unless these priorities are set in advance, and understood by everyone concerned, the analyst cannot make trade-off decisions without incessant consultation. The analyst is also apt to be surprised by the reaction of users or management to aspects of performance that have been ignored. If the support and use of the system crosses organizational boundaries, you might need a written service-level agreement between the providers and the users to ensure that there is a clear common understanding of the performance objectives and priorities.
In general, the performance of a given workload is determined by the availability and speed of one or two critical system resources. The analyst must identify those resources correctly or risk falling into an endless trial-and-error operation.
Systems have both real and logical resources. Critical real resources are generally easier to identify, because more system performance tools are available to assess the utilization of real resources. The real resources that most often affect performance are as follows:
Logical resources are less readily identified. Logical resources are generally programming abstractions that partition real resources. The partitioning is done to share and manage the real resource.
Some examples of real resources and the logical resources built on them are as follows:
It is important to be aware of logical resources as well as real resources. Threads can be blocked by a lack of logical resources just as for a lack of real resources, and expanding the underlying real resource does not necessarily ensure that additional logical resources will be created. For example, consider the NFS block I/O daemon (biod, see Tuning for NFS Performance). A biod daemon on the client is required to handle each pending NFS remote I/O request. The number of biod daemons therefore limits the number of NFS I/O operations that can be in progress simultaneously. When a shortage of biod daemons exists, system instrumentation may indicate that the CPU and communications links are used only slightly. You may have the false impression that your system is underused (and slow), when in fact you have a shortage of biod daemons that is constraining the rest of the resources. A biod daemon uses processor cycles and memory, but you cannot fix this problem simply by adding real memory or converting to a faster CPU. The solution is to create more of the logical resource (biod daemons).
Logical resources and bottlenecks can be created inadvertently during application development. A method of passing data or controlling a device may, in effect, create a logical resource. When such resources are created by accident, there are generally no tools to monitor their use and no interface to control their allocation. Their existence may not be appreciated until a specific performance problem highlights their importance.
Consider minimizing the workload's critical-resource requirements at three levels, as discussed below.
The decision to use one resource over another should be done consciously and with specific goals in mind. An example of a resource choice during application development would be a trade-off of increased memory consumption for reduced CPU consumption. A common system configuration decision that demonstrates resource choice is whether to place files locally on an individual workstation or remotely on a server.
For locally developed applications, the programs can be reviewed for ways to perform the same function more efficiently or to remove unnecessary function. At a system-management level, low-priority workloads that are contending for the critical resource can be moved to other systems, run at other times, or controlled with the Workload Manager.
Because workloads require multiple system resources to run, take advantage of the fact that the resources are separate and can be consumed in parallel. For example, the operating system read-ahead algorithm detects the fact that a program is accessing a file sequentially and schedules additional sequential reads to be done in parallel with the application's processing of the previous data. Parallelism applies to system management as well. For example, if an application accesses two or more files at the same time, adding an additional disk drive might improve the disk-I/O rate if the files that are accessed at the same time are placed on different drives.
The operating system provides a number of ways to prioritize activities. Some, such as disk pacing, are set at the system level. Others, such as process priority, can be set by individual users to reflect the importance they attach to a specific task.
A truism of performance analysis is that there is always a next bottleneck. Reducing the use of one resource means that another resource limits throughput or response time. Suppose, for example, we have a system in which the utilization levels are as follows:
CPU: 90% Disk: 70% Memory 60%
This workload is CPU-bound. If we successfully tune the workload so that the CPU load is reduced from 90 to 45 percent, we might expect a two-fold improvement in performance. Unfortunately, the workload is now I/O-limited, with utilizations of approximately the following:
CPU: 45% Disk: 90% Memory 60%
The improved CPU utilization allows the programs to submit disk requests sooner, but then we hit the limit imposed by the disk drive's capacity. The performance improvement is perhaps 30 percent instead of the 100 percent we had envisioned.
There is always a new critical resource. The important question is whether we have met the performance objectives with the resources at hand.
If, after all of the preceding approaches have been exhausted, the performance of the system still does not meet its objectives, the critical resource must be enhanced or expanded. If the critical resource is logical and the underlying real resource is adequate, the logical resource can be expanded at no additional cost. If the critical resource is real, the analyst must investigate some additional questions: