<p>A parallel file system is essential for executing large workloads via large-scale computing resources such as national leadership computers; however, a primary cause of performance degradation arises from I/O operation. In the scientific workloads, there is often a significant variance in I/O performance, and the information needed to determine the characteristics of the tasks is limited at the time of job submission. These characteristics make it very difficult to analyze and predict I/O performance. Nevertheless, accurate analysis of the I/O behavior of applications based on multidimensional logs greatly aids in overcoming these challenges. In this paper, we examine in detail the relationship between applications and I/O behavior on the basis of multiple logs from jobs executed on the national leadership computer <Emphasis FontCategory="NonProportional">Nurion</Emphasis>, which is operated by the Korea Institute of Science and Technology Information (KISTI). Moreover, we analyze which key features best reflect the characteristics of the applications and propose methods to identify applications on the basis of this analysis. In practice, the proposed classification method can eliminate the tedious tasks of users having to manually input the types of applications. In addition, it is useful for administrators to restrict the types and number of concurrently running applications for various reasons. To verify this potential, our experimental results include the accuracy of applying four algorithms widely used in multiclass classification on the basis of key features detected through feature analysis, along with a confusion matrix.</p>