From Yefu's notes
Jump to: navigation, search

HPCA07的Concurrent Direct Network Access for Virtual Machine Monitors提高VM中的网络效率。

Energy Management for Commercial Servers Server节能的几种办法

ECOSystem: Managing Energy as a First Class Operating System Resource详细说了CPU, 磁盘和网卡

Operating-system directed power reduction详述了在操作系统中实现节能的办法。大约是按需给硬盘和NIC降速。

Conserving Disk Energy in Network Servers比较了给磁盘节能的四种方法,并指出最好的方法是使用两个异速磁盘。


VEE'09的Dynamic memory balancing for virtual machines在Xen3.2上实现了动态分配内存的算法。ASPLOS'09上的PowerNap: Eliminating Server Idle Power降低了server的静态能耗,其中对各个部件动静能耗、响应时间的描述,并指出server在很多时间都是idle的。CARVE: A Cognitive Agent for Resource Value Estimation用一种机器学习的方法在虚拟平台中分配资源(以内存为例),旨在获得正的收益。该论文采用了VIOS。 Memory-aware Dynamic Voltage and Frequency Prediction for Portable Devices Performance Directed Energy Management for Main Memory and Disks

Software thermal management of dram memory for multicore systems., SIGMETRICS'08

VEE'09 VEE'08 VEE'07

Memory Buddies: Exploiting Page Sharing for Smart Colocation in Virtualized Data CentersThis paper proposes algorithms to consolidate VMs to servers such that more pages of memories can be shared. Experiments are based on VMWare ESX

Post-copy based live virtual machine migration using adaptive pre-paging and dynamic self-ballooningThis paper propose a post-copy based live migration algorithm which transfers memory pages on demand.

Netchannel: a VMM-level mechanism for continuous, transparentdevice access during VM migration


Power-aware QoS management in Web servers大约是用了admission control,所以既控制了power也控制了QoS. 该文说server farm最节能的办法是exact load balancing.

Performance Modeling of an Apache Web Server with Bursty Arrival Traffic,in In Proceedings of International Conference on Internet Computing (IC}, This paper provides a model of Apache which is able to deal with bursty traffic.

A Reinforcement Learning Approach to Online Web Systems Auto-configurationChange parameters like MaxClient Joint optimization of idle and cooling power in data centers while maintaining response time 里面用了G1/G/m的模型 Adaptive Control of Multi-Tiered Web Application Using Queueing Predictor用了M/GI/1/PS模型+adaptive control.


Entropy: a Consolidation Manager for Clusters实现了一个最小化机器数的migration算法。Coordinated management of power usage and runtime performance,还有pManager。

Efficient Resource Provisioning in Compute Clouds via VM Multiplexing首先说VM都峰谷交替,有数据。在每个VM violate capacity的概率是\beta_i的情况下,推出了整机的capacity要满足的公式。


Network-Aware Dynamic Voltage and Frequency Scaling通过DVFS来控制Network的能量,但我感觉只有网络发送方才行。


Energy-Efficient Real-Time Heterogeneous Server ClustersWorkload dispatching+Server on/off Self-organized server farms for energy savings

Optimal power allocation in server farms, SIGMETRICS'09This paper studies in a server farm with a workload distributor at the front, which power configuration is optimal. e.g., under a data center level power budget, run few servers at high speed and keep the rest shut down.

Self-adaptive admission control policies for resource-sharing systems SIGMETRICS'09

Managing server energy and operational costs in hosting centers

A distributed control framework for performance management of virtualized computing environments: some preliminary results那个look ahead control的扩展。里面有模型,考虑了response time, queue length, request rate, CPU share;但是看不出来这个模型是怎么回事。


Reducing Energy Consumption of Disk Storage Using Power-Aware Cache Management




Real-Time Dynamic Power Management through Device Forbidden Regions



Just-in-time server provisioning using virtual machine standby and request prediction


Self‐Adaptive and Self‐Configured CPU Resource Provisioning for Virtualized Servers Using Kalman Filters This paper allocates CPU resources to VMs to guarantee CPU utilization using Kalman Filter.

vManage: Loosely Coupled Platform and Virtualization Management in Data CentersThis paper coordinates virtulization-management (VM mapping) and platform management (SLA, server-level power budget) in a way of trial-and-error. They assume the CDF of any resource requirement (CPU, mem, etc.) is known in prior.

VCONF: A Reinforcement Learning Approach to Virtual Machines Auto-configuration This paper is not available on line yet.

On the Use of Fuzzy Modeling in Virtualized Data Center Management


A comparison of software and hardware techniques for x86 virtualization

Geiger: monitoring the buffer cache in a virtual machine environment

Temporal search: detecting hidden malware timebombs with virtual machines.


Weighted fair sharing for dynamic virtual clusters (2 pages)


(HPCA'08)C-Oracle: Predictive Thermal Management for Data Centers C-Oracle predicts thermal effect of different actions such as load distribution and DVFS. It is a model-based prediction based on utilization of components (CPU, disk). C-Freon uses C-oracle, request distribution, admission control, server shutdown to perform DTM. C-Freon uses DVFS+admission control. Their experiments are on real servers.

(HPCA'07, PSU)Modeling and Managing Thermal Profiles of Rack-mounted Servers with ThermoStat This paper designs a thermal modeling tools for servers and server racks by using Fluid Dynamics. Their examples argue that the temperatures of servers in the same rack affect each other. Just like HotSpot at server level. The model is validated in a real server.

(ECRTS'07, U. of Pitt.)Thermal Faults Modeling using a RC model with an Application to Web Farms This paper provides a power model of processor considering failure events, e.g., a fan failure. The model is validated on a real server with an extra sensor. As an application, they design an algorithm to deal with thermal emergencies for clusters in case of thermal failure using open loop search.

[ICAC'06]Weatherman: Automated, Online, and Predictive. Thermal Mapping and Management for Data Centers. This work uses AI to predict temperature of every point in a room given a thermal topology, then uses heuristics to find a workload placement configuration to minimise cooling costs by using coordinate space search (heuristics). Results are from simulator Flovent.

[Distributed and Parallel Databases, 2007]Thermo-Fluids Provisioning of a High Performance High Density Data Center (short paper). This paper provides model to depict the thermo-fluids in data centers. They use Flovent simulator and validated in a real datacenter.

[COMSWARE 2007]Software Architecture for Dynamic Thermal Management in Datacenters. This paper propose thermal aware job scheduling considering ambient and on-board temperatures. Experiments are finished in a real datacenter.

[ASPLOS'06] "Mercury and Freon: Temperature Emulation and Management for Server Systems” Mercury is a server-level temperature simulator. Freon uses load balancing and admission control to avoid overheating in clusters. Freon-EC extended Freon to save energy by shutting down servers.

(IEEE Internet Computing, 2005, by HP) Balance of Power: Dynamic Thermal Management for Internet Data Centers This paper aims at balancing the temperature of servers in a row of a cluster, or in a data center. The actuator is the power of every server. The algorithm is equivalent to a P-controller. This paper is simulation only.

(USENIX'05) “Making Scheduling Cool: Temperature-Aware Workload Placement in Data Centers”. This paper extends the HP paper in IEEE Internet Computing. By workload placement, they actually mean power allocation to servers. This paper aims at minimising the power consumed by air conditioning. To achieve this, they maximise the temperature of the cold air given by the air conditioning device by allocating power to servers. Simulation results are based on Flovent simulator. They have two algorithms. ZBD is extended from the HP paper. MinHR is open loop search.

(ICAC'07, U. of Toronto)Adaptive Learning of Metric Correlations for Temperature-Aware Database Provisioning This paper aims at controlling delay in DB data centers with temperature awareness to avoid hot spot. They use machine learning (SVM).


{Power and Performance Management of Virtualized Computing Environments Via Lookahead Control}{Kusic_icac2008}. This paper dispatches requests among VMs to optimize the performance. The actuator they use is the dispatch ratio. The goal is to maximize a revenue function. The benchmark they used is IBM Trade6.

\item \textbf{Application performance management in virtualized server environments}\cite{Khanna_Migration}. This paper performs VM migration to maximize a function which contains the response time, migration cost(CPU and memory) and the cost to add new servers. The benchmark used is Trade3. The workload generator is Websphere Workload Simulator. \item \textbf{Profiling, prediction, and capping of power consumption in consolidated environments}\cite{choi_power_prediction} This paper provides a prediction algorithm to predict the average and peak power of a VM though the off line data. They use TPC-W and SPEC CPU 2000. \item \textbf{Efficient Power Management of Heterogeneous Soft Real-Time Clusters}\cite{wang_dispatch}. This paper uses workload dispatching among a cluster of servers ti minimize the power consumption of the whole server The idle servers are turned off to save power. They use simulation only. 1000 Islands This paper uses three levels of controllers to (1) guarantee the response time and (2) migrate VMs to satisfy resource needs. They do not aim at minimizing power. The benchmark is Apache and a self-programmed CGI program.

\item \textbf{Algorithms for non-uniform size data placement on parallel disks}\cite{Kashyap_disks}. This paper aims at migrating media files among several disks with bandwidth constraints. The problem is the same as ours mathematically. We aims at migrating VMs among servers with CPU constraints.

\item \textbf{Delivering Energy Proportionality with Non Energy-Proportional Systems – Optimizing the Ensemble}\cite{tolia}. This paper aims at turning common server systems into energy-proportional systems, i.e., the energy is proportional to its usage. The policy they use is DVFS+VM migration. In addition, they study the energy proportionality of the cooling device.

\item \textbf{Feedback Driven QoS-Aware Power Budgeting for Virtualized Servers}\cite{ripal_power_bid}. This paper controls the power of the physical server to a budget while controlling QoS of VMs. It first decides the total cap of all VMs based on the power budget based on a PID controller, than allocates this total cap among VMs based on a bidding like algorithm.



Guest-Aware Priority-Based Virtual Machine Scheduling for Highly Consolidated Server This paper schedules the VMs based on the priority of the processes inside the VMs. This method increases the response time to IO events by 5~22%. Our work is different from theirs in that 1) we provide performance guarantee for every VM, 2) their algorithm relies on the fact that every VM would be assigning the priorities to the processes properly in a nice way, hence face the risk that a VM will oversubscribe the CPU time by increasing priorities to all its processes unnecessarily.

Xen and Co.: Communication-aware CPU Scheduling for Consolidated Xen-based Hosting Platforms This paper advocates an algorithm preferring giving CPU to the VM with the most number of network packages to minimize the delay of the packages caused by virtulization.

Scheduling I/O in Virtual Machine Monitors Traditional VM schedulars only consider the sharing of CPU with no concern of I/O performance. Though an I/O-intensive may receive enough total CPU time, it may not receive the CPU resource at the right time. This paper boots the I/O performance of VMs by giving priority to the VMs with pending I/O requests.

Trace file

M. Arlitt and T. Jin. A workload characterization study of the 1998 world cup web site. Network, IEEE, 14(3):30{37, May/Jun 2000.



Adaptive model predictive control for co-ordination of compression and friction brakes in heavy duty vehicles

Adaptive Model Predictive Control of Multivariable Time-varying Systems