The current diverse and wide range of computing moves towards the cloud and de- mands high performance in low latency and high throughput. Facebook reported that 3.3 billion people monthly and 2.6 billion people daily use their data centers over the network. Many emerging user-facing applications require strict control over the stor- age latency’s tail to provide a quality user experience. The low-latency requirement triggers the ongoing replacement of hard drives (HDDs) by solid-state drives (SSDs) in the enterprise, enabling much higher performance and lower end-to-end storage latencies. It becomes more challenging to ensure low latency while maintaining the device’s endurance ratings. We address this challenge in the following ways: 1. Enhance the overall storage system’s performance and maintain the SSD endurance using emerging Non-volatile memory (ENVM) technology. 2. Implement deterministic la- tency in the storage path for latency-sensitive applications. 3. Provide low-latency and differentiated services when write-intensive workloads are present in a shared environment. We have proposed the performance and endurance-centric mechanisms to evaluate the tradeoffs between performance and endurance. In the first approach, our goal is to achieve low storage latency and a long lifetime of the SSD simultane- ously, even for a write-heavy workload. Incorporating a significantly smaller amount of ENVM with SSD as a cache helps to achieve the said goal.SSDs using the NVMe (Non-Volatile Memory Express) interface can achieve low latency as the interface provides several advanced features. The second approach has iii explored such features to control the storage tail latency in a distributed environment. The ”predictable latency mode (PLM)” advanced feature helps to achieve determinis- tic storage latency. SSDs need to perform many background management operations to deal with the underlying flash technology traits, the most time-consuming ones be- ing garbage collection and wear leveling. The latency requirement of latency-sensitive applications violates when the I/O requests fall behind such management activities. PLM leverages SSD controllers to perform the background operations during a win- dow, called a ”non-deterministic window (NDWin)”. Whereas during the ”determin- istic window (DTWin)”, applications will experience no such operations. We have extended this feature in the distributed environment and showed how it helps achieve low storage latency when the proposed ”PLM coordinator (PLMC)” is incorporated. In a shared environment with write-intensive workloads present, result in latency peak for Read IO. Moreover, it is required to provide differentiated services with multiple QoS classes present in the workload mixture. We have extended the PLM concept on hybrid storage to realize the deterministic latency for tight tail-controlled appli- cations and assure differentiated services among multiple QoS applications. Since nearly all of the storage access in a data center is over the network, an end-to-end path consists of three components: The host component, Network component, and Storage Component. For latency-sensitive applications, the overall tail latency needs to consider all these components. In a NAS (Network Attached Storage) architecture, it is worth studying the QoS class aware services present at the different components to provide an overall low request-response latency. Therefore, it helps future research to embrace the gaps that have not been considered yet. / Computer and Information Science

The current diverse and wide range of computing moves towards the cloud and de- mands high performance in low latency and high throughput. Facebook reported that 3.3 billion people monthly and 2.6 billion people daily use their data centers over the network. Many emerging user-facing applications require strict control over the stor- age latency’s tail to provide a quality user experience. The low-latency requirement triggers the ongoing replacement of hard drives (HDDs) by solid-state drives (SSDs) in the enterprise, enabling much higher performance and lower end-to-end storage latencies. It becomes more challenging to ensure low latency while maintaining the device’s endurance ratings. We address this challenge in the following ways: 1. Enhance the overall storage system’s performance and maintain the SSD endurance using emerging Non-volatile memory (ENVM) technology. 2. Implement deterministic la- tency in the storage path for latency-sensitive applications. 3. Provide low-latency and differentiated services when write-intensive workloads are present in a shared environment. We have proposed the performance and endurance-centric mechanisms to evaluate the tradeoffs between performance and endurance. In the first approach, our goal is to achieve low storage latency and a long lifetime of the SSD simultane- ously, even for a write-heavy workload. Incorporating a significantly smaller amount of ENVM with SSD as a cache helps to achieve the said goal.SSDs using the NVMe (Non-Volatile Memory Express) interface can achieve low latency as the interface provides several advanced features. The second approach has iii explored such features to control the storage tail latency in a distributed environment. The ”predictable latency mode (PLM)” advanced feature helps to achieve determinis- tic storage latency. SSDs need to perform many background management operations to deal with the underlying flash technology traits, the most time-consuming ones be- ing garbage collection and wear leveling. The latency requirement of latency-sensitive applications violates when the I/O requests fall behind such management activities. PLM leverages SSD controllers to perform the background operations during a win- dow, called a ”non-deterministic window (NDWin)”. Whereas during the ”determin- istic window (DTWin)”, applications will experience no such operations. We have extended this feature in the distributed environment and showed how it helps achieve low storage latency when the proposed ”PLM coordinator (PLMC)” is incorporated. In a shared environment with write-intensive workloads present, result in latency peak for Read IO. Moreover, it is required to provide differentiated services with multiple QoS classes present in the workload mixture. We have extended the PLM concept on hybrid storage to realize the deterministic latency for tight tail-controlled appli- cations and assure differentiated services among multiple QoS applications. Since nearly all of the storage access in a data center is over the network, an end-to-end path consists of three components: The host component, Network component, and Storage Component. For latency-sensitive applications, the overall tail latency needs to consider all these components. In a NAS (Network Attached Storage) architecture, it is worth studying the QoS class aware services present at the different components to provide an overall low request-response latency. Therefore, it helps future research to embrace the gaps that have not been considered yet. / Computer and Information Science