User-Level I/O Accelerations For High-Performance Deep Learning Applications