Optimized Allocation of Distributed Applications Across Local Area Networks

Enterprise-wide distributed computing systems are inherently stochastic and the performance management of distributed applications running on the top of DCS is a complex and computationally hard task. In this paper, we define the enterprise distributed application and provide queueing analysis on the relationship between the application level performance parameters and the network and computer system parameters. We also provide a binary integer programming (BIP) model and a case study on how to allocate application components across an enterprise local area network. 1. Distributed Application and Network Model We model the network as nodes and connections. The nodes include computers and other network switching apparatus. A connection links a pair of nodes, and can be direct through one link, or indirect through multiple links and other nodes. Nodes can be used to distribute application components, but only some dedicated server machines are used for this purpose. We call these dedicated server machines distributing nodes. For all distributing nodes, we assume that a connection or route exists between every pair of nodes. Two nodes may be connected through one or more switching nodes and links. A link is described by its bandwidth or data rate. We define following parameters in our model: 1) Node memory size is described as MEMSi for node i, 2) Average disk I/O time is described as IOTi for node i, 3) A link l is described by its bandwidth Bl, 4) A router r is described by its router latency RLr, 5) An application component a requires MEMa size of node memory at run time, 6) The average CPU time for component a on node i is described as Sai, 7) Ka denotes the average number of visit to disk I/O for component a to execute a transaction job, 8) Current CPU utilization U ′ i (CPU) for node i, 9) Current disk utilization U ′ i (disk) for node i, 10) Current memory usage MEM ′ i for node i, 11) Current load γ ′ l in packet per time unit for link l, 12) Current utilization U ′ r for router r. We now discuss the mapping of response time to network and system parameters. We model a distributed application system as an open queueing network. These include node queues and connection queues (note a connection may include multiple links and routers). For simplicity, we model each service provided by nodes and connections as a M/M/1 queue, and we call each node or connection a service facility. We assume the whole application system to be a markovian system, which means the distributions of the inter-arrival times and service times are exponential distributions (Poisson) that exhibit the markov (memoryless) property. Within the system, service facilities may have inputs from multiple facilities and outputs to multiple facilities. We assume all input and output processes are Poisson processes and that merged or split Poisson processes are still Poisson processes. As an example, we modeled the distributed application (with process flows) described in [1] with the Jackson Network in Figure 1.