Date of Award

Summer 2019

Project Type

Dissertation

Program or Major

Computer Science

Degree Name

Doctor of Philosophy

First Advisor

Robert D Russell

Second Advisor

Radim Bartos

Third Advisor

Philip Hatcher

Abstract

As computer networking has evolved and the available throughput has increased, the efficiency of the network software stack has become increasingly important. This is because the latency introduced by software has gone from insignificant, compared to historically poor network performance, to the largest component of latency for a modern local-area network. Currently, the vast majority of code that accesses the hardware is part of the kernel, because the kernel is responsible for ensuring that user applications do not interfere with each other when accessing the hardware. Remote Direct Memory Access~(RDMA) provides a solution for applications to perform direct data transfers over the network without requiring context switches into the kernel, but relies instead on specialized hardware interfaces to handle the virtual address mappings and transport protocols. This more intelligent hardware allows for direct control from the userspace application, eliminating the cost of context switches into the kernel. This in turn reduces the overall latency of message transfers.

Just like networking, storage is currently undergoing a similar evolution. For most of the recent history of computing, the most common durable storage mechanism has been mechanical hard disk drives, which can only be accessed at block level and have high latency compared to the software drivers used to access the data. However, the introduction of solid state disks~(SSDs) based on Flash significantly decreased the latency, as there are no mechanical parts that need to move to access the data. Upcoming non-volatile memory solutions reduce this latency even further, and even allow byte-level access to the storage medium. Thus, just like with networking, software drivers become the bottleneck and we look for solutions to bypass the kernel to improve the efficiency of direct userspace access to storage.

This thesis offers two contributions as part of a solution to these problems. The first part introduces urdma, a software RDMA driver which leverages the Data Plane Development Kit (DPDK) to perform network data transfers in userspace without specialized RDMA interface hardware. The second part examines remote locking protocols, which are required for synchronization in distributed storage systems. We define an RDMA locking mechanism referred to as Verbs Offload Locking Technology (VOLT), which allows acquisition of a remote lock object without any CPU usage by the target node. This offloading allows VOLT to be used with disaggregated memory servers that have limited onboard CPU resources, while also lowering the application overhead for remote locking. Finally, we define a bytecode framework using enhanced Berkeley Packet Filter (eBPF) bytecode for extending the capabilities of an RDMA-capable network interface card (NIC) with new operations, and show how this can be used to implement our remote locking operation.

Download

Share

COinS