WallMon : Interactive distributed monitoring of process-level resource usage on display and compute clusters

Abstract

To achieve low overhead, traditional cluster monitoring systems sample data at low frequencies and with coarse granularity. However, interactive monitoring requires frequent (up to 60 Hz) sampling of fine-grained data and visualization tools that can explore and display data in near real-time. This makes traditional cluster monitoring systems unsuited for interactive monitoring of distributed cluster applications, as they fail to capture short-duration events, making understanding the performance relationship between processes on the same or different nodes difficult. To address this issue, WallMon was developed, a tool for interactive visual exploration of performance behaviors in distributed systems. For gathering of data, WallMon is centered around an abstraction of collectors and handlers; collectors gathers data of interest, such as CPU and memory usage, and forwards it to handlers in a push-based fashion, while handlers take action upon the data. WallMon captures and visualizes data for every process on every node, as well as overall node statistics. Data is visualized using a technique inspired by the concept of information flocking. WallMon's design is based on the client-server model, and it is extensible through a module system that encapsulates functionality specific to monitoring (collectors) and visualization (handlers). A set of experiments have been carried out on a cluster of 29 nodes with 180 processes per node. Performance results show 7% (of 100) CPU usage at 64 Hz sampling rate when performing process-level monitoring with WallMon. Using WallMon's interactive visualization, we have observed interesting patterns in different parallel and distributed systems, such as unexpected ratio of user- and kernel-level execution among processes in a particular distributed system.