PIOMAT: Parallel File I/O System for Many-cores Extracting High Throughput Potential
A parallel file I/O system for many-core coprocessors extracting their high throughput potential is designed and implemented in this work. The many-core is attached to the host processor via PCIe bus. A file I/O server is running on the host and it is composed of kernel threads dedicated for processing file I/O requests from the many-core and accesses the file system running on the host. The congestion on the host file I/O server or file system may occur in the case of massive amount of requests coming from the many-core due to the limited throughput of the host and the bandwidth of PCIe bus.
For the purpose of solving this problem, data caching structure called as File Space, which utilizes x86 virtual address space and can show fine scalability and throughput, is proposed and developed on the many-core. The evaluation result verifies the appropriateness of adopting data caching strategy on the many-core and shows that the aggregative bandwidth can scale for parallel writing in the case of 48 threads and the performance stands out from that of Linux provided by Intel in both cases of using NFS mounting the host file system and employing local file system on many-core.