Article Preview
Top1. Introduction
In the realms of Big Data analysis, Medical analysis, network analysis, and image analysis, inadequate multivariate high dimensional data poses the biggest challenge (Chapman et al., 2000; Grosu & Chronopoulos, 2005). In general, these spheres contain data, mostly every sort of variables that are usually inadequate. In this backdrop, Multiple Imputations is considered offering the best possible practice that handles insufficiency in data (Ambler et al., 2007; Dempster et al., 1977; Dempster et al., 1977; Bu et al., 2016). However, the currently available multiple imputation algorithms suffer from issues such as high time complexity, often lacking the good properties that distributed and shared processing possess. Hence, we find them unsuitable to process data in a high-dimensional data ecosystem. Further, in view of the rapid movement of systems that run this task towards exascale, which demands communication and computation patterns needing high programmability (Li et al., 2014; Wang et al., 2015). However, such systems are quite complex, which the current communication models find difficult to locate and productively utilize for computation-communication overlap as High-Performance Computing (HPC) usually suffer from performance as well as programming related issues (Bu et al., 2016; Feng & Balaji, 2009; Humenay et al., 2007). The available communication models are found lacking close energy with multi-threaded programming models that happen in a rough or risky scenario in which communication and multi-threaded components of applications are synchronized. In fact, it is rather tough to program distributed memory systems having huge quantities of parallelism in every node. The available distributed memory models are intended to achieve scalability and communication. But in some particular applications, they are found unsuitable as programming models that leverage overt parallelism. But, shared memory task models are ideally suited to exploit overt parallelism. This paper proposes a set of Distributed algorithms for improving the efficiency of an MI task. A Cloud computing ecosystem is found appropriate for such a system that contains homogeneous as well heterogeneous resources. Moreover, multi-cluster resources have the ability to enhance computing capacity seen while executing a task among workers in data of higher dimensions. This ecosystem needs a suitable resource management model for having a more effective as well as genuine multiple imputation task strategy. This paper suggests novel algorithms that include features such as, (i) Serial Regression MI algorithm for running MI task in a distributed system; (ii) a shared-memory based MI algorithm with load balancing for running MI task; computation and estimating work efficiency parameters with standards such as NRMSE, Bias, Running Time and Speed (Bu et al., 2016). Cloud platforms offer highly reliable, available and better download capacity minus the hassles that researchers face while dealing with on-site assets or logistics, mostly providing newer scopes for exploration. The suggested algorithm betters the process in which data is handled as well as cloud computing resource are utilized. Further, it helps submission of complex tasks in a transparent and convenient manner (Amir et al., 2000). It suggests carrying out large scale experiments for comparing the suggested one with at least a couple of typical missing data imputation algorithms like Distributed Serial Regression Multiple Imputation using Socket (DSMIS), Distributed Serial Regression Multiple Imputation using fork (DSMIF) and Distributed Serial Regression Multiple Imputation using fork with Load Balance (DSMIFLB). The results showed the suggested algorithm attaining improved imputation precision by taking much less time compared to the existing algorithms as far as imputing high-dimensional data in multivariate distribution is concerned. The rest of this paper is divided into the following sections with section 2 provides Literature Study, and section 3 covers the required background study. Further, section 4 provides a description of the suggested approach, while, in section 5, evaluated the experimental results with relation to related work, and finally, section 6 concludes the work by providing a few observations as well as insights into future work.