Phil C. Cayton

Sensor-Based Fast Thermal Evaluation Model For Energy Efficient High-Performance Datacenters

In this work, we propose an abstract heat flow model which uses temperature information from onboard and ambient sensors, characterizes hot air recirculation based on these information, and accelerates the thermal evaluation process for high performance datacenters. This is critical to minimize energy costs, optimize computing resources, and maximize computation capability of the datacenters. Given a workload and thermal profile, obtained from various distributed sensors, we predict the resulting temperature distribution in a fast and accurate manner taking into account the recirculation characterization of a datacenter topology. Simulation results confirm our hypothesis that heat recirculation can be characterized as cross interference in our abstract heat flow model. Moreover, fast thermal evaluation based on cross interference can be used in online thermal management to predict temperature distribution in real-time.

Thermal-Aware Task Scheduling to Minimize Energy Usage of Blade Server Based Datacenters

Blade severs are being increasingly deployed in modern datacenters due to their high performance/cost ratio and compact size. In this study, we document our work on blade server based datacenter thermal management. Our goal is to minimize the total energy costs (usage) of datacenter operation while providing a reasonable thermal environment for their reliable operation. Due to special characteristics of blade servers, we argue that previously proposed power-oriented schemes are ineffective for blade server-based datacenters and that task-oriented scheduling is a more practicable approach since the contribution to the total energy cost from cooling and computing systems varies according to the utilization rates. CFD simulations are used to evaluate scheduling results of three different task scheduling algorithms: uniform outlet profile (UOP), minimal computing energy (MCE), and uniform task (UT), under four different blade-server energy consumption models: discretenonoptimal (DNO), discreteoptimal (DO), analognonoptimal (ANO), and analogoptimal (AO). Simulation results show that the MCE algorithm, in most cases, results in a minimal total energy cost - a conclusion that differs from the findings of previous research. UOP performs better than UT at low datacenter utilization rates, whereas UT outperforms UOP at high utilization rates

Software Architecture for Dynamic Thermal Management in Datacenters

Minimizing the energy cost and improving thermal performance of power-limited datacenters, deploying large computing clusters, are the key issues towards optimizing their computing resources and maximally exploiting the computation capabilities. In this paper, we develop a unique merger between the physical infrastructure and resource management functions of a cluster management system to take a holistic view of datacenter management, and make global (at the level of a datacenter) thermal-aware job scheduling decisions. A software architecture is presented in this regard and implemented in a fully operational computational cluster in the ASU datacenter. The proposed architecture develops a feedback-control loop, by combining information from ambient and on-board sensors with the node allocation and job scheduling mechanisms, for managing the system load depending on the thermal distribution in the datacenter.

Efficient direct user level sockets for an Intel/spl reg/ Xeon/spl trade/ processor based TCP on-load engine

Intel Labs has continued development of the embedded transport acceleration (ETA) software prototype that uses one of the Intel/spl reg/ Xeon/spl trade/ processors in a multi-processor server as a packet processing engine (PPE) that is closely tied to the servers core CPU and memory complex. We have further developed the prototype to provide support for user-level, asynchronous interface for sockets. The direct user socket interface (DUSI) allows user-level applications to interface directly to the PPE using familiar socket commands and semantics. The prototype runs in an asymmetric multiprocessing mode, in that the PPE does not run as a general computing resource for the host operating system. We describe the prototype software architecture, the DUSI application interface, and detail our measurement and analysis of some micro-benchmarks. In particular, we measure throughput for transactions and end-to-end latency as the key metrics for the analysis.