A Hybrid Supervised Learning Approach for Detection and Mitigation of Job Failure with Virtual Machines in Distributed Environments

Abstract

Distributed data processing techniques are very popular nowadays due to high data generation from various resources. To increase work learning outcomes and to reduce consumption, modern massive computational systems divide jobs into several smaller tasks that perform in parallel. Nevertheless, responding with straggler processes, which are sluggish running processes that rise the total response time, is a typical performance issue in such platforms. In this paper, we proposed the detection of struggler nodes in a large distributed environment using a hybrid machine learning technique. Initially, the data has been collected from numerous virtual machine network logs. The entire data set has various fields such as Virtual Machine ID, CPU load, memory load, bandwidth utilization, etc. Memory utilization an input to the proposed system is collected from the garbage collection log files where the memory consumption on each VM and its timestamp is recorded. This is the most efficient way to get the memory consumption in web/desktop applications. Similarly, the CPU, I/O and bandwidth utilization is grabbed from the process monitoring functionality and SAR (System Activity Report) utility from the respective VM boxes. This data set is useful to identify weather-specific virtual machine is heated up or not. In this approach, we proposed three conventional machine learning algorithms and a hybrid machine learning algorithm for the identification of node status. Main purpose of the proposed system is to identify the slow performing node in an efficient way to prevent the other nodes from failures. This can provide effective load balancing and low response time for task execution from available VM’s in distributed cloud environments. To create its training program, several extractions of features approaches were used. TF-IDF, correlational co-occurrence, and density-based features have been mined from the whole data set. With extensive experimental analysis, we evaluate our system with our proposed classification algorithm. As a result, the system produces higher classification accuracy of 94.5% over the traditional machine learning classifiers. If the proposed system is tested against the data set fields, memory load and CPU load on the homogenous machine configurations, we see more efficiency while detecting the underperforming node than the heterogenous machine configurations.