End-to-End Considerations in Unification of High-Performance IO

The performance of modern distributed storage and computing frameworks considerably depends on the IO performance of the many storage and network devices involved. Fortunately, these IO devices have undergone a rapid transformation in the past decade and are now capable of delivering multi-Gigabits/sec bandwidths and ultra-low IO latencies. However, in contrast to IO devices, the performance improvements of single CPU have stalled in the same time period. Hence, the traditional notion of a single fast CPU connected to multiple slow devices no longer holds. Yet, IO stacks are still designed to optimize the CPU time by executing multiple services and routines on a fast CPU while a slow IO operation is in progress. This situation has led to a CPU-IO performance gap, where the CPU’s inability to keep up with the execution of thick software stacks and OS routines during a fast IO operation on high-performance network and storage devices limits the performance delivered to data-crunching applications. Multiple research efforts from industry as well as academia have been launched to improve this situation by reducing the hardware and software overheads by providing better IO interfaces, efficiently managing IO resources, and leveraging manycore CPUs for IO processing. However, these efforts exclusively either target the network or the storage stack but not the combination of both. In this thesis, we address this performance gap and advocate to take a holistic approach towards managing resources, data flows, and devices (network or storage) to form end-to-end data flows in a distributed setting. We first quantify the software and OS overhead in IO operations and and argue to reduce it by building upon the high-performance networking principle. The general philosophy of the principle is to recognize and separate the slow control setup from the fast data access path, and involve CPU/OS in the former only selectively in managerial roles. In the thesis framework, we extend the separation philosophy from networks to storage devices by identifying common themes in the evolution of their software stacks. After identifying common high-performance IO properties, we make a case to unify the network and storage stacks. We then design and build a proof of concept FlashNet, a unified software IO stack that uses highperformance networking abstractions and semantics to access remote storage. In accordance with the original separation philosophy, FlashNet allows the allocation and translation of both