Towards maximizing the efficiency of datacenter power and server infrastructure
Open Access
- Author:
- Narayanan, Iyswarya
- Graduate Program:
- Computer Science and Engineering
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- September 11, 2019
- Committee Members:
- Anand Sivasubramaniam, Dissertation Advisor/Co-Advisor
Anand Sivasubramaniam, Committee Chair/Co-Chair
Chitaranjan Das, Committee Member
Bhuvan Urgaonkar, Committee Member
Hosam K. Fathy, Outside Member
Timothy Zhu, Committee Member
Chitaranjan Das, Program Head/Chair - Keywords:
- datacenter
power management
resource management
capacity planning
energy storage in a datacenter
datacenters - Abstract:
- Applications increasingly rely on the cloud infrastructure that spans several datacenter facilities for their computing needs. Datacenter operators expend billions of dollars in provisioning and operating their physical infrastructure. Hence, it is critical for the datacenter operators to extract maximum performance from their physical infrastructure. While there are prior works on datacenter infrastructure optimization, they are insufficient in the context of modern datacenters due to two main reasons: First, prior works on optimizing datacenter infrastructure focus primarily for a single datacenter whereas the physical infrastructure of today's large datacenter operators is spread across multiple facilities located in different geographies. This leads to inefficiencies when provisioning capacity across several facilities to meet stringent latency and availability needs of user-facing applications. Second, the effectiveness of datacenter operations closely depends on the efficiency of individual datacenters. While prior works primarily focus on improving efficiency of server infrastructure within a datacenter, modern datacenters are power constrained because of cost and environmental concerns. Hence, it is critical to maximize the efficiency of both power infrastructure as well as server infrastructure. Towards addressing these concerns, this thesis takes a holistic approach to optimize the physical infrastructure of datacenters. This thesis first explores how to reduce inefficiencies when planning datacenter capacity across multiple facilities. Capacity considerations for a geo-distributed infrastructure do not decompose into individual datacenter capacity planning. Heterogeneous availability and costs of datacenter facilities, non-uniform spatial distribution of clients, and interdependence between latency and availability constraints of the applications make it non-trivial to provision the right capacity at each datacenter. This thesis develops a geo-distributed capacity planning framework to capture the key factors that influence capacity, ranging from application demand patterns, latency and availability requirements, datacenter cost-availability trade-offs, and data replication overheads. Next, the thesis focuses on improving the efficiency of individual datacenters by effectively utilizing resources in both power and server infrastructure. Most of today's servers, with numerous CPU cores and other plentiful direct resources, co-locate workloads to improve server utilization. However, power is an equally important indirect resource in a server that is shared between the co-located applications, for which they can contend, especially when power budgets are tight. We refer to this as a ``power struggle". We first explore this problem in the context of a private-cloud cluster which is provisioned for a primary latency-critical application, but also admits secondary best-effort applications to improve utilization during its off-peak periods. Here, even if the power capacity is tuned for the peak load of the primary application, co-locating another application with it during its off-peak period can lead to overshooting of the power capacity. Therefore, to extract maximum returns on datacenter infrastructure investments one needs to jointly handle power and server resources. Towards that goal, our solution draws on principles from economics theory to reason about resource demands in power constrained environments and provides answers to the when, where, and what questions pertaining to co-location. In addition to tackling the aforementioned power struggles due to static infrastructure capacity limits, this thesis further studies this problem when a datacenter imposes dynamic power budgets on individual servers to regulate its overall power consumption. While there is a considerable amount of prior effort on server power capping, they are largely oblivious to power as an indirectly shared resource. This indirect resource exhibits a unique set of properties -- spatial non-multiplexing, non-convex, space and time-shifting -- which have not been explicitly tackled so far. We propose power management policies that explicitly consider these properties to effectively mediate power struggles under stringent power constraints. Finally, this thesis explores tapping into energy storage devices within the power hierarchy of a datacenter to engage in several emerging power modulation requirements (e.g. peak shaving, renewable integration, power regulation) with diverse characteristics. Different energy storage technologies exist, including little explored technologies such as flow batteries, that offer different performance characteristics in cost, size, and environmental impact. While prior works in datacenter energy storage have considered one of usage aspects, technology, performance metric (typically cost), the whole three-dimensional space is little explored. To understand this design space, this thesis presents joint characterization of energy storage usage scenarios on their provisioning and operating fronts, under ideal and realistic energy storage technologies, and quantifies the impact on datacenter performance.