Architectural Enhancements for Data Transport in Datacenter Systems

Datacenter systems run myriad applications, which frequently communicate with each other and/or Input/Output (I/O) devices—including network adapters, storage devices, and accelerators. Due to the growing speed of I/O devices and the emergence of microservice-based programming models, the I/O software stacks have become a critical factor in end-to-end communication performance. As such, I/O software stacks have been evolving rapidly in recent years. Datacenters rely on fast, efficient “Software Data Planes”, which orchestrate data transfer between applications and I/O devices. The goal of this dissertation is to enhance the performance, efficiency, and scalability of software data planes by diagnosing their existing issues and addressing them through hardware-software solutions. In the first step, I characterize challenges of modern software data planes, which bypass the operating system kernel to avoid associated overheads. Since traditional interrupts and system calls cannot be delivered to user code without kernel assistance, kernel-bypass data planes use spinning cores on I/O queues to identify work/data arrival. Spin-polling obviously wastes CPU cycles on checking empty queues; however, I show that it entails even more drawbacks: (1) Full-tilt spinning cores perform more (useless) polling work when there is less work pending in the queues. (2) Spin-polling scales poorly with the number of polled queues due to processor cache capacity constraints, especially when traffic is unbalanced. (3) Spin-polling also scales poorly with the number of cores due to the overhead of polling and operation rate limits. (4) Whereas shared queues can mitigate load imbalance and head-of-line blocking, synchronization overheads of spinning on them limit their potential benefits. Next, I propose a notification accelerator, dubbed HyperPlane, which replaces spin-polling in software data planes. Design principles of HyperPlane are: (1) not iterating on empty I/O queues to find work/data in ready ones, (2) blocking/halting when all queues are empty rather than spinning fruitlessly, and (3) allowing multiple cores to efficiently monitor a shared set of queues. These principles lead to queue scalability, work proportionality, and enjoying theoretical merits of shared queues. HyperPlane is realized with a programming model front-end and a hardware microarchitecture back-end. Evaluation of HyperPlane shows its significant advantage in terms of throughput, average/tail latency, and energy efficiency over a state-of-the-art spin-polling-based software data plane, with very small power and area overheads. Finally, I focus on the data transfer aspect in software data planes. Cache misses incurred by accessing I/O data are a major bottleneck in software data planes. Despite considerable efforts put into delivering I/O data directly to the last-level cache, some access latency is still exposed. Cores cannot prefetch such data to nearer caches in today's systems because of the complex access pattern of data buffers and the lack of an appropriate notification mechanism that can trigger the prefetch operations. As such, I propose HyperData, a data transfer accelerator based on targeted prefetching. HyperData prefetches exact (rather than predicted) data buffers (or a required subset to avoid cache pollution) to the L1 cache of the consumer core at the right time. Prefetching can be done for both core-peripheral and core-core communications. HyperData's prefetcher is programmable and supports various queue formats—namely, direct (regular), indirect (Virtio), and multi-consumer queues. I show that with a minor overhead, HyperData effectively hides data access latency in software data planes, thereby improving both application- and system-level performance and efficiency.PhDComputer Science & EngineeringUniversity of Michigan, Horace H. Rackham School of Graduate Studieshttp://deepblue.lib.umich.edu/bitstream/2027.42/169826/1/hosseing_1.pd