Browsing by Subject "Datacenter"

Heterogeneous design presents an opportunity to improve energy efficiency but raises a challenge in management. Whereas prior work separates the two, we coordinate heterogeneous design and management. We present a market-based resource allocation mechanism that navigates the performance and power trade-offs of heterogeneous architectures. Given this management framework, we explore a design space of heterogeneous processors and show a 12x reduction in response time violations when equipping a datacenter with three processor types over a homogeneous system that consumes the same power. To better understand trade-offs in large heterogeneous design spaces, we explore dozens of design strategies and present a risk taxonomy that classifies the reasons why a deployed system may underperform relative to design targets. We propose design strategies that explicitly mitigate risk, such as a strategy that minimizes the coefficient of variation in performance. In our experiments, we find that risk-aware design accounts for more than 70% of the strategies that produce systems with the best service quality. We also present a new datacenter management mechanism that fairly allocates processors to latency-sensitive applications. Tasks express value for performance using sophisticated piecewise-linear utility functions. With fairness in market allocations, we show how datacenters can mitigate envy amongst latency-sensitive users. We quantify the price of fairness and detail efficiency-fairness trade-offs. Finally, we extend the market to fairly allocate heterogeneous processors.

As demands for users’ applications’ data increase, the world’s computing platforms are moving towards more capable machines – servers and warehouse-scale datacenters. Diverse users share datacenters for complex computation and compete for shared resources. In some systems, such as public clouds where users pay for reserved hardware, management policies pursue performance goals. In contrast, private systems consist of users who voluntarily combine their resources and subscribe to a common management policy. These users reserve the right to opt-out from shared systems if resources are managed poorly. The system management framework needs to ensure fairness among strategic users, encouraging users to participate while guaranteeing individual performance and preserving the system’s performance. Microeconomic models are well suited for studying individual behavior and the allocation of scarce resources. In this thesis, we present three pieces of work on task colocation, resource allocation, and task scheduling problems to demonstrate the effectiveness of a microeconomic approach.

Colocating applications on shared hardware (i.e., chip-multiprocessors) improves server utilization but introduces resource contention into the memory subsystem. In the first work, we design a colocation framework based on cooperative game theory to manage shared resource contention. Our framework uses a recommendation system to predict individual applications preferences for colocated tasks. It then uses these predictions to drive novel colocation mechanisms to guarantee user fairness and preserve system performance. Attractive system outcomes encourage strategic users to participate in the datacenter.

Processor allocations are inefficient when they are based on static reservations because reservations are often conservative; users rarely know their application’s needs across time, especially when applications have complex phase behavior. In the second work, we propose a fast, lightweight performance prediction framework to help users capture their phase behaviors in parallel applications. We design a dynamic and distributed core allocation framework so that users can trade resources for better efficiency based on predicted performance. Our management framework provides efficient allocations and game-theoretic fairness guarantees. In the last work, we characterize applications’ sensitivity to non-uniform memory access (NUMA) in big memory servers. We develop performance and energy models for communication costs in a blade server. We use this model to perform case studies on NUMA-aware scheduling policies and task queue management. Our parameterized models lay the foundation for the coordinated design of scheduling policies and hardware configurations. This method can be further used to design locality-aware schedulers with microeconomic models, e.g., dynamic pricing strategies for city parking.