TACKLING PERFORMANCE AND SECURITY ISSUES FOR CLOUD STORAGE SYSTEMS

Building data-intensive applications and emerging computing paradigm (e.g., Machine Learning (ML), Artificial Intelligence (AI), Internet of Things (IoT) in cloud computing environments is becoming a norm, given the many advantages in scalability, reliability, security and performance. However, under rapid changes in applications, system middleware and underlying storage device, service providers are facing new challenges to deliver performance and security isolation in the context of shared resources among multiple tenants. The gap between the decades-old storage abstraction and modern storage device keeps widening, calling for software/hardware co-designs to approach more effective performance and security protocols. This dissertation rethinks the storage subsystem from device-level to system-level and proposes new designs at different levels to tackle performance and security issues for cloud storage systems. In the first part, we present an event-based SSD (Solid State Drive) simulator that models modern protocols, firmware and storage backend in detail. The proposed simulator can capture the nuances of SSD internal states under various I/O workloads, which help researchers understand the impact of various SSD designs and workload characteristics on end-to-end performance. In the second part, we study the security challenges of shared in-storage computing infrastructures. Many cloud providers offer isolation at multiple levels to secure data and instance, however, security measures in emerging in-storage computing infrastructures are not studied. We first investigate the attacks that could be conducted by offloaded in-storage programs in a multi-tenancy cloud environment. To defend against these attacks, we build a lightweight Trusted Execution Environment, IceClave to enable security isolation between in-storage programs and internal flash management functions. We show that while enforcing security isolation in the SSD controller with minimal hardware cost, IceClave still keeps the performance benefit of in-storage computing by delivering up to 2.4x better performance than the conventional host-based trusted computing approach. In the third part, we investigate the performance interference problem caused by other tenants' I/O flows. We demonstrate that I/O resource sharing can often lead to performance degradation and instability. The block device abstraction fails to expose SSD parallelism and pass application requirements. To this end, we propose a software/hardware co-design to enforce performance isolation by bridging the semantic gap. Our design can significantly improve QoS (Quality of Service) by reducing throughput penalties and tail latency spikes. Lastly, we explore more effective I/O control to address contention in the storage software stack. We illustrate that the state-of-the-art resource control mechanism, Linux cgroups is insufficient for controlling I/O resources. Inappropriate cgroup configurations may even hurt the performance of co-located workloads under memory intensive scenarios. We add kernel support for limiting page cache usage per cgroup and achieving I/O proportionality