Xiao Qin's Research

BUD: A Buffer-Disk Architecture for Energy Conservation in Parallel Disk Systems

Download BUD Annual Report PDF

Findings

An Energy-Efficient Framework for Large-Scale Parallel Storage Systems
A Simulation Framework for Energy-efficient Data Grids
An Energy-Efficient Scheduling Algorithm Using Dynamic Voltage Scaling for Parallel Applications on Clusters
Load-Balancing Strategies for Energy-Efficient Parallel Storage Systems with Buffer Disks
Sacrificing Reliability for Energy Saving: Is It Worthwhile for Disk Arrays
Load-Balancing Strategies for Energy-Efficient Parallel Storage Systems with Buffer Disks
Energy-Efficient Prefetching for Parallel I/O Systems with Buffer Disks
Improving Reliability and Energy Efficiency of Disk Systems via Utilization Control
Energy Conservation for Real-Time Disk Systems with I/O Burstiness
An Adaptive Energy-Conserving Strategy for Parallel Disk Systems
DARAW: A New Write Buffer to Improve Parallel I/O Energy-Efficiency
MICRO: A Multi-level Caching-based Reconstruction Optimization for Mobile Storage Systems
PEARL: Performance, Energy, and Reliability Balanced Dynamic Data Redistribution for Next Generation Disk Arrays
HyBUD: An Energy-Efficient Architecture for Hybrid Parallel Disk Systems
Energy-Aware Prefetching for Parallel Disk Systems
DORA: A Dynamic File Assignment Strategy with Replication
Collaboration-Oriented Data Recovery for Mobile Disk Arrays
A File Assignment Strategy Independent of Workload Characteristic Assumptions
References

Handling read operations is kind of simple and straightforward in the BUD framework. When a read request arrives, the controller first searches the RAM buffer. If the data is still in the RAM buffer then the data is immediately sent back to the requester. Otherwise, the controller will do a seek operation in the buffer disks. If the required data can not be found in the buffer disks, a miss message will be sent back to controller and the controller will send a read request to the corresponding data disk and finally the data will be transferred to the requester by the data disk. Using this policy, the read performance should be similar to or sometimes better than that of traditional disk because most of the requests will be sent to the data disk and reading from RAM or buffer disks seldom occurs in real applications.

Our preliminary results consist of first developing a simulator which meets all project specifications and running this simulator with a trace to get some preliminary results [Zong et al., 2007b]. So far the simulator completed all the main functions that are necessary in order to model our distributed system. That is the program takes data from a trace. Then the program moves data to appropriate virtual disks that use disksim to derive there timing information. These virtual disks use a simple model to calculate total energy. Finally, both timing and energy data are reported to the user in the form of the two respective totals. Our results from the simulator consist of two parallel disk systems. To simulate with these two systems we used a simple trace that came with disksim. The first system that was simulated was a simple disk system which is used today by many storage systems. It is basically a RAID 1 system consisting of 31 disks. This is basically a baseline system in which to compare our results. The second simulated disk system is similar to section 3 with 6 buffer disks each one acting as a buffer for a group of approximately 5 disks.

The results from running these systems showed that both took 25.55 min. Table below shows the energy consumption of the disk system without employing buffer disks and the disk system using buffer disks to conserve energy. Specifically, the traditional system without buffer disks consumed 189279.78 J (0.05 KWH) with all disks starting from off and being turned on when needed. In contrast, the parallel disk system with buffer disks only used 117345.99 J (0.03 KWH), which is substantially less. This resulted in overall power savings of 38%. Our results have shown substantial gains can be made by using buffer disks for very specific workload conditions. In order to further develop our storage algorithm we need to test our results in many more simulations and with more varied workloads. The simulator also needs to check with analytic calculations and the code needs to include the transition time and power from switching from mode to mode.

Simulated Disk System Energy Consumption Disk system without buffer disks 189279.78 J Disk system with buffer disks 117345.99 J Energy saving 71933.79 J Energy consumption reduced by 38%

Simulated Disk System	Energy Consumption
Disk system without buffer disks	189279.78 J
Disk system with buffer disks	117345.99 J
Energy saving	71933.79 J
Energy consumption reduced by	38%

[Zong et al., 2007c] High performance data grids are increasingly becoming popular platforms to support data-intensive applications. Reducing high energy consumption caused by data grids is a challenging issue. Most previous studies in grid computing focused on performance and reliability without taking energy conservation into account. As such, designing energy-efficient data grid systems became highly desirable. In this project, we proposed a framework to simulate energy-efficient data grids. We presented an approach to integrate energy-aware allocation strategies into energy-efficient data grids. Our framework aims at simulating a data grid that can conserve energy for data-intensive applications running on data grids.

A data grid can be envisioned as a complicated distributed system, which consists of the following four major layers: application layer, middleware layer, resource layer and network layer. Fig. 1 depicts the four layers and their relations for data grids. It is worth noting that the simulation framework is constructed in accordance with the system architecture outlined as follows:

Data grids are complex multivariate environments, which are made up of numerous grid entities that need to be automatically managed. In order to make coherent and coordinated use of ubiquitous and heterogeneous data and storage resources, resource management is a centerpiece in the simulation framework. In this section, we present a framework to simulate data grids that are energy efficient in nature. In general, a data grid should energy efficiently handle two important components in the system: storage resources and data-intensive jobs. A data grid has to ad-dress the following issues. First, it is of paramount impor-tance to find available storage resources within a short pe-riod of time. Second, it must allocate data-intensive jobs to available resources. We proposed a simulation framework for data grids:

In this framework, the global level scheduler (or grid level scheduler) coordinates mul-tiple local scheduling or helps to select the most appropri-ate resources for a job among different possible resources. Typically, a global level scheduler itself has no direct control over computing and storage resources. Therefore, the global level scheduler has to communicate with and appropriately trigger several local level schedulers to complete data-intensive tasks submitted by users. Those local level schedulers either control resources directly or have certain access to their local resources. The global level scheduler is also responsible for collaborating with other important supportive middleware like information services, communication services, and reliability control modules.

[Zong et al., 2007c] As long as the grid scheduling module collected all the information of currently available computing and storage resources, it can judiciously choose target recourses based on its scheduling policy and allocate the tasks analyzed by the task analyzer to these chosen resources for parallel execution. We designed the job scheduling flow in the simulation framework:

During the process of execution, the results collector will periodically check the randomly returned sub results and transfer these sub results to Grid level scheduler. The scheduler in the framework passes the latest information to all tasks, which can guarantee that the tasks with dependency could immediately be executed once they get the necessary sub results.

[Ruan et al., 2007] In the past decade cluster computing platforms have been widely applied to support a variety of scientific and commercial applications, many of which are parallel in nature. However, scheduling parallel applications on large scale clusters is technically challenging due to significant communication latencies and high energy consumption. As such, shortening schedule length and conserving energy consumption are two major concerns in designing economical and environmentally friendly clusters. In this study, we proposed an energy-efficient scheduling algorithm (TADVS) using the dynamic voltage scaling technique to provide significant energy savings for clusters. The TADVS algorithm aims at judiciously leveraging processor idle times to lower processor voltages (i.e., the dynamic voltage scaling technique or DVS), thereby reducing energy consumption experienced by parallel applications running on clusters. Reducing processor voltages, however, can inevitably lead to increased execution times of parallel task. The salient feature of the TADVS algorithm is to tackle this problem by exploiting tasks precedence constraints. Thus, TADVS applies the DVS technique to parallel tasks followed by idle processor times to conserve energy consumption without increasing schedule lengths of parallel applications. Experimental results clearly show that the TADVS algorithm is conducive to reducing energy dissipation in large-scale clusters without adversely affecting system performance.

In the first set of experiments, we varied CCR from 0.1 to 1 to examine the performance impacts of communication intensity on our TADVS scheduling strategy. This year, we evaluate the performance of TADVS algorithm by comparing the traditional NDS scheme: