Workload Management in WebLogic Server 9.0
by Naresh Revanuru
03/07/2006
Abstract
Workload management is the term used to describe how client-generated work is accepted and handled in an application server. BEA WebLogic Server 9.0 introduces new concepts for workload management, not available in previous releases. These concepts replace execute queues as defined in earlier releases, and include notions of work prioritization, thread pool management, and overload protection. This article describes how workload management is handled in WebLogic Server 9.0, and how it differs from previous releases.
Introduction
Prior to WebLogic Server 9.0, customers had to configure execute queues with a fixed thread count and a pending work queue length. However, dealing directly with low-level kernel attributes such as the number of threads has problems. Here are some of the difficulties that administrators face:
- Given an application, it is difficult to determine the exact number of threads required. It takes a lot of trial and error to come up with a thread count that gives reasonably good performance.
- Customers are forced to create separate execute queues to achieve work prioritization. For example, consider a scenario with two applications deployed on the same server instance. The administrator needs to make sure that both the applications get a fair share of the thread resources at any given point in time irrespective of the arrival rate. The only way to achieve this guarantee is to configure two separate execute queues with their own dedicated threads for each application.
- Sometimes execute queues need to be configured for some requests within the same application because they interact with a slow-moving back end and have the potential to block faster requests from executing because the arrival rate of the slow-moving requests is much greater. This is sometimes referred to as the convoy effect. Prior to WebLogic Server 9.0, creating dedicated thread pools was the only solution to all such problems.
- Dedicated execute queues have to be configured to guarantee ordering and to make sure that there is at most one thread picking up work for a certain class of requests.
- Each execute queue creates its own thread pool. The number of threads in the server is the sum of thread counts of all execute queues in the server. Execute queues are configured for maximum capacity and performance. This means that there could be a large number of threads in the server that are underutilized but that still must be monitored. Imagine a thread dump with more than 100 threads! It is difficult to diagnose problems.
With WebLogic Server 9.0 administrators can move away from configuring dedicated execute queues and start describing application requirements in a language they understand. Here are some of the salient features:
- Applications can request their own fair share of thread resources by using the fair-share parameter. This tells the WebLogic Server kernel to allocate threads to the two applications proportionately, irrespective of their arrival pattern. In fact, this is done by default without any configuration.
- Administrators no longer have to configure thread counts, and the optimal thread count is determined by the server automatically based on the overall throughput measurements.
- It is possible to set constraints to enforce minimum and maximum concurrency without creating dedicated thread pools.
- Unlike execute queues, all WorkManagers share a common thread pool and a priority-based queue. The size of the thread pool is determined automatically by the kernel and resized as needed. WorkManagers (described on the next page) become very lightweight, and customers can create WorkManagers without worrying about the size of the thread pool. Thread dumps look much cleaner with fewer threads.
- In the new model, it is possible to specify different service-level agreements (SLAs) such as fair shares or response-time goals for the same servlet invocation depending on the user associated with the invocation. For example, it is possible to give higher fair share to "platinum" users than to "evaluation" users for the same servlet or EJB invocation.
I will describe more about these features in the following sections.
New Thread Pool Implementation
WebLogic Server 9.0 has a single thread pool for requests from all applications. Similarly, all pending work is enqueued in a common priority-based queue. The thread count is automatically tuned to achieve maximum overall throughput. Priority of the requests is dynamic and computed internally to meet the stated goals. Administrators state goals simply, using application-level parameters like fair-share, response time goals.
In earlier releases, each servlet or RMI request was associated with a dispatch policy that mapped to an execute queue. Requests without an explicit dispatch policy use the server-wide default execute queue. In WebLogic Server 9.0, requests are still associated with a dispatch policy but are mapped to a WorkManager instead of to an execute queue. Note that the concept of WorkManager described here is completely different from the Timer and WorkManager specification described on page 5.
Requests without an explicit dispatch policy use the default WorkManager of the application. This means that each application has its own default WorkManager that is not shared with other applications. This distinction is important to note. Execute queues are always global whereas WorkManagers are always application scoped. Even WorkManagers defined globally in the console are application scoped during runtime. This means that each application gets into own runtime instance that is distinct from others, but all of them share the same characteristics like fair-share goals. This is done to track work at the application layer and to provide capabilities like graceful suspension of individual applications.
As mentioned earlier, each servlet or RMI request is associated with a WorkManager. By default, all requests are associated with the application default WorkManager. The dispatch-policy element can be used to associate a request with a specific WorkManager either defined within an application scope or globally at the server level. I will provide examples on how to use the dispatch policy in a later section.
Thread Count Self-tuning
One of the major differences between execute queues and the new thread scheduling model is that the thread count does not need to be set. In earlier releases, customers defined new thread pools and configured their size to avoid deadlocks and provide differentiated service. It is quite difficult to determine the exact number of threads needed in production to achieve optimal throughput and avoid deadlocks. WebLogic Server 9.0 is self-tuned, dynamically adjusting the number of threads to avoid deadlocks and achieve optimal throughput subject to concurrency constraints. It also meets objectives for differentiated service. These objectives are stated as fair shares and response-time goals as explained in the next section.
The self-tuning thread pool monitors the overall throughput every two seconds and uses the collected data to determine if thread count needs to change. Present thread count, the measured throughput, and the past history is taken into account by the algorithm to determine if the thread count needs to increase or decrease, and new threads are automatically added to the pool or removed, as needed.