Service Level Definitions and Interactions

This section starts with a high level view of service level management and defines a terminology that is based on existing practices.

Computer systems are used to provide a service to end users. System and application vendors provide a range of components that can be used to construct a service. System managers are responsible for the quality of this service. A service must be available when it is needed and must have acceptable performance characteristics.

Service level management is the process by which information technology (IT) infrastructure is planned, designed, and implemented to provide the levels of functionality, performance, and availability required to meet business or organizational demands.

Service level management involves interactions between end users, system managers, vendors and computer systems. A common way to capture some of these interactions is with a service level agreement (SLA) between the system managers and the end users. Often, many additional interactions and assumptions are not captured formally.

Service level management interactions are shown in FIGURE 2-1. Each interaction consists of a service definition combined with a workload definition. There are many kinds of service definitions and many views of the workload. The processes involved in Service Level Management include creating service and workload definitions and translating from one definition to another.

The workload definition includes a schedule of the work that is run at different times of the day (for example, daytime interactive use, overnight batch, backup, and maintenance periods). For each period, the workload mix is defined in terms of applications, transactions, numbers of users, and work rates.

The service level definition includes availability and performance for service classes that map to key applications and transactions. Availability is specified as uptime over a period of time and is often expressed as a percentage (for example 99.95 percent per month). Performance may be specified as response time for interactive transactions or throughput for batch transactions.

System Level Requirements

System managers first establish a workload definition and the service level requirements. The requirements are communicated to vendors, who respond by proposing a system that meets these requirements.

Sizing Estimates

Vendors measure the system performance using generic benchmarks. They may also work with system managers to define a customer-specific benchmark test. Vendors provide a sizing estimate based on the service level requirements and workload definition. The basis of the sizing estimate can be a published benchmark performance. In some cases, the measured performance on a customer-defined benchmark is used as the basis of a sizing estimate. Vendors provide reliability data for system components. They can also provide availability and performance guarantees for production systems with a defined workload (at a price). Vendors cannot provide unqualified guarantees because typically, many application and environmental dependencies are outside their control.

Service Level Agreements

System managers and end users negotiate an SLA that establishes a user-oriented view of the workload mix and the service levels required. This may take the form: 95th percentile response time of under two seconds for the new-order transaction with up to 600 users online during the peak hour. It is important to specify the workload (in this case the number of users at the peak period), both to provide bounds for what the system is expected to do and to be precise about the measurement interval. Performance measures averaged over shorter intervals will have higher variance and higher peaks.

The agreed-upon service levels could be too demanding or too lax. The system may be quite usable and working well, but still failing an overly demanding service level agreement. It could also be too slow when performing an operation that is not specified in the SLA or whose agreed-to service level is too lax. The involved parties must agree to a continuous process of updating and refining the SLA.

Real User Experiences

The actual service levels experienced by users with a real workload are subjective measures that are very hard to capture. Often problems occur that affect parts of the system not covered explicitly by the service level agreement, or the workload varies from that defined in the service level agreement. One of the biggest challenges in performance management is to obtain measurements that have a good correlation with the real user experience.

Service Level Measurements

The real service levels cannot always be captured directly, but the measurements taken are believed to be representative of the real user experience. These measurements are then compared against the service level agreement to determine whether a problem exists. For example, suppose downtime during the interactive shift is measured and reported. A problem could occur in the network between the users and the application system that causes poor service levels from the end-user point of view but not from the system point of view. It is much easier to measure service levels inside a backend server system than at the user interface on the client system, but it is important to be aware of the limitations of such measurements. A transaction may take place over a wide variety of systems. An order for goods will affect systems inside and outside the company and across application boundaries. This problem must be carefully considered when the service level agreement is made and when the service level measurements are being defined.

Policies and Controls

System managers create policies that direct the resources of the computer system to maintain service levels according to the workload definition specified in those policies. A policy workload definition is closely related to the service level agreement workload definition, but may be modified to satisfy operational constraints. It is translated into terms that map oto system features. Example policies include:

• A maximum of 600 interactive users of the order entry application at any time.

• Order entry application has a 60 percent share of CPU, 30 percent share of network, and 40 percent share of memory.

• If new-order response time is worse than its target, steal resources from other workloads that are overachieving.

The policy is only as effective as the measurements available to it. If the wrong things are being measured, the policy will be ineffective. The policy can control resources directly or indirectly. For example, direct control on CPU time and network bandwidth usage might be used to implement indirect control on disk I/O rates by slowing or stopping a process.

Capacity Planning and Exception Reporting

The measured workload and service levels should be analyzed to extract trends. A capacity planning process can then be used to predict future scenarios and determine action plans to tune or upgrade systems, modify the service level agreement, and proactively avoid service problems.

In cases where the measured workload from the users exceeds the agreed-upon workload definition or where the measured service level falls short of the defined level, an exception report is produced.

Accounting and Chargeback

The accrued usage of resources by each user or workload may be accumulated into an accounting system so that projects can be charged in proportion to the resources they consume.