Controlling and Scheduling Parallel I/O in Multi-application Environments

As clusters usage grows, a lot of scientific applications (biology, climatology, nuclear physics ...) have undergone rewrites to harness the extra CPU and extra storage provided. These demanding software, besides handling huge amounts of data with peculiar parallel I/O access patterns, are run on clusters, environments where concurrency between those applications occurs. Several propositions have been made to manage both the intensive parallel I/O applications and the cluster constraints. Nevertheless, available Parallel File Systems or Parallel I/O Libraries are based on specific API's, which limit portability and require good knowledge of their internal mechanisms to get good performances. Moreover, Parallel I/O Libraries are usually focused on running only one application without taking into account the load that the other ones generate on the cluster. This paper presents a new strategy to handle parallel I/O in a multi-application and distributed environment. Our framework detects parallel I/O accesses without inter-processes synchronization mechanisms and uses a simple interface based on the classic UNIX system calls (creat/open/read/write/close). In addition, we analyze two scheduling strategies to improve global performances and provide fairness between applications. Early experiments have given promising results and have shown that using such approaches may lead to better performances as well as improvements of quality of service