ENERGY EFFICIENT QUALITY OF SERVICE PROVISIONING IN DISAGGREGATED STOARGE SYSTEM

Abstract

With the emergence of high-performance storage devices and emerging remote access protocols such as NVMe over Fabrics (NVMe-oF), system throughput has increased several times. However, at the same time, aggregate throughput and uncertainty of traffic intensity can easily clog the network links and cause multiple QoS violations. Thus provisioning Quality of Services (QoS) becomes more challenging in such disaggregated storage systems, as it has to rely on both the storage and the network for performance. Besides that, energy management plays a crucial role in QoS provisioning, and provisioning needs to be managed cost-effectively (e.g., minimize energy consumption). Though most works mainly focus on the internal component of the storage and host servers, the network becomes a power-hungry resource as the legacy interconnect networks are being replaced with high-speed links (100 Gbps) that consumes several times more power than the conventional one. Also, at the same time, the increasing speeds generally result in much lower network utilization, which allows energy saving. From that motivation, we leverage the Low Power Idle (LPI) feature of Ethernet and propose an intelligent routing mechanism in order to maximize low power (or "sleep") opportunities for the network interfaces while avoiding link congestion. In testing out our mechanisms, we have enhanced the popular ns3 package for network modeling. Our intelligent routing mechanism serves the purpose of better performance and power tradeoff in networking components unless there is congestion in the storage server itself. To address that, we develop novel mechanisms for dynamically deciding when to move storage chunks (storage access unit) or alter the number of active chunk copies to alleviate congestion during high traffic episodes and enable traffic consolidation (and hence network energy savings) during low traffic periods. Increasing the number of active chunk copies (replication) or migrating storage chunks give the opportunity of performance enhancement during high load period and powersaving during the low period of activity. Using extensive simulations with modified ns3 network simulator, we provide deep insights into the migration vs. replication tradeoff. Though network congestion can be alleviated, tremendous throughput from different storage servers can still congest the TOR ⇐⇒ Host link or TOR ⇐⇒ Storage link in a shared environment. In that scenario, service differentiation is necessary, as different applications have different QoS requirements. In that context, we propose a QoS aware transport layer solution, which offers QoS differentiation without any compromise in overall system throughput.