Improving I/O performance for High Performance Computing with Application Forwarding Layer

In the use of High Performance Computing platforms, I/O performance has become an important limiting factor. For regular users, after a supercomputer is put into use, I/O performance optimization based on the physical structure is usually difficult to achieve. In this paper, we show that this can be achieved through modifying the application layer. Considering that a supercomputer has many computing nodes, we present an approach that uses dedicated computing nodes to process the requests from different applications to optimize the I/O performance. The dedicated computing nodes perform aggregation and reorder the operations before sending the requests to the storage system. We call the role played by these computing nodes in this process the logical I/O forwarding layer which corresponds to the I/O forwarding layer of the physical structure. Our experiments show that this approach is feasible and provides room for further improvement.