Modeling Sun Cluster Availability
Maximizing the availability of computer systems and services is becoming the primary focus in IT environments today 1,2. Coupled with a desire to use lower cost solutions than traditional fault tolerant systems, this focus has resulted in an exponential growth in the deployment of clusters as highly available platforms. Clusters typically use redundant off-the-shelf hardware components together with specialized software to provide highly available services at a lower cost.
Evaluating the availability offered by a clustering solution can be done in a variety of ways, ranging from analyzing the availability offered by the system to measuring it, both in pre- and post-deployment phases. Typically, analyzing complex software systems such as clusters is significantly harder than measuring the corresponding metrics, but at the same time, the availability model derived from analysis is more accurate. Software modeling also becomes increasingly difficult as more software components are added to the system. Evaluating software reliability and availability has been studied both in academia and in industry for the past few decades3. While much progress has been made in this research, an accurate assessment of general software availability remains an intractable task.
The inherent difficulty of accurately evaluating software availability is primarily due to the exponential number of ways in which its various software components can interact in different usage scenarios. This is exacerbated by the following facts:
a significant percentage of these interactions are typically not completely understood due to the components having been created by different groups.
Software components keep changing at a fast pace that defies expedient analysis of their interactions with other potentially changing components.
However, there is a huge demand to evaluate the availability of a clustering solution for several reasons:
Customers can use this metric as a yardstick to differentiate among clustering solutions offered by different vendors.
Customers can use this metric to determine the availability of their applications running on the cluster.
Software designers can use this metric to help understand what part of the software needs redesign or changes to improve the availability of the system.
This article reports on an effort underway at Sun to model Sun Cluster availability for specific configurations that include the data service layer. It describes the proposed methodology that employs a hybrid approach to tackle the inherently NP-complete4 nature of this task, with the aim of delivering a reasonably accurate availability model.
The remainder of this article is organized as following sections:
"RAScad Modeling" describes a top-level availability model of a two-node Sun Cluster configuration using Sun's RAScad tool5. In the context of this article, the term node refers to a server machine.
"Proposed Methodology" discusses the salient components of the methodology proposed for modeling Sun Cluster availability.
"The Stack" details the hardware and software configuration (referred to as a stack in this article) under investigation, and gives some preliminary results measured for this stack. These values are applied to the RAScad model, and the results of differential analysis are presented in this section. This analysis yields a set of best practices for any enterprise.
"Best Practices" discusses a set of best practices for any enterprise.
"Conclusion" presents concluding remarks.
The starting point of modeling a Sun Cluster system is to build a top-level system availability behavioral model with RAScad. RAScad is a Sun internal reliability, availability, and serviceability architecture modeling and analysis tool for use in computer system design and development phases. It highlights the different variables that contribute to system availability, as well as the degree to which they affect it.
FIGURE 1 shows the RAScad Markov model for a two-node Sun Cluster stack for any fault that causes one of the nodes to go down. This fault could be an operating system panic or any hardware fault causing the node to fail. The data service is a scalable service, meaning that each of the two nodes is hosting the service actively. A node going down results in reconfigurations of the clustering framework as well as of the service related components in this stack.
FIGURE 1 RAScad Availability Model for a Two-node Sun Cluster Stack
In FIGURE 1 a vertex in the graph represents the system state, and an edge represents the failure rate of the transition between the source and sink states of that edge. A value of 1 marks a state if it represents the service being up or available, otherwise a value of 0 marks the state.
Under steady-state, the cluster is in state Both_Up where, as the name implies, both nodes are functional, servicing client requests. The cluster can transition out of this state if one of the nodes goes down; the rate of either of the two nodes dying is 2/MTBF, where MTBF is the Mean Time Between Failures for a node. This event takes the cluster into the Recovery state, where it remains unavailable while the clustering and the service components reconfigure. The reconfiguration completes successfully with a probability p, and takes Recovery_Time to do so. A successful reconfiguration takes the cluster into an available state, One_Up.
In state One_Up, the surviving node services client requests, but at a higher load level since clients that were being serviced by two nodes are now being serviced by just one. This causes the failure rate of the node to increase by a factor of a, which results in a corresponding decrease in the value of the MTBF of the node.
The time taken to repair the failed node is MTTR_1, where MTTR stands for the Mean Time to Repair a node. After being repaired, it is booted back into the cluster. When it joins, the cluster goes into an unavailable state Node_Rejoin, while the joiner node gets refreshed with the existing cluster state. The average time taken for this is Node_Rejoin_Time, after which the cluster enters the original available state, Both_Up.
The node in state One_Up can go down with a failure rate of (1+a)/MTBF, and the cluster would enter state Both_Down, where both nodes of the cluster have gone down.
Following the second outgoing path from state Recovery, an unsuccessful reconfiguration takes the cluster to an unavailable state, Both_Down. Assuming a repair policy where both nodes are repaired at the same time, the average time to repair and boot back the two nodes is MTTR_2, after which the cluster goes back to the original, available state, Both_Up.
Once the key parameters in this model are identified as described previously, RAScad can be used to compute steady state system availability, system yearly downtime and other related useful metrics. Furthermore, a differential analysis of the effect of the various parameters on the availability of the system for specified value ranges provides very useful information not only about what factors to focus on in product design, development, and testing, but also to tailor the availability assessment efforts to an appropriate subset of scenarios. Perhaps most importantly from a customer's perspective, by clearly pinpointing the outages that must be avoided in order to maintain high availability of a system, this differential analysis helps formulate a set of rules or best practices that should be adopted in any Sun Cluster environment to maximize availability in any enterprise. In "Best Practices," this model is used for computing the various availability metrics and to perform a differential analysis for the stack under discussion.