Debugging Inconsistent Network Performance
High-performance network interconnects, such as Infiniband, are common in high performance computing environments. Recently, my colleagues, and I ran a series of experiments on the Cooley system at the Argonne Leadership Computing Facility. For one set of experiments, the network performance was consistent and fast, but for the others we periodically had poor performance. Up until this point in my work, I had not directly used or configured network libraries below this layer....