A key requirement of the cloud platform is the reasonable deployment of its large-scale virtual machine infrastructure. The mapping relation between the virtual node and the physical node determines the specific resource distribution strategy and reliability of the virtual machine deployment. Resource distribution strategy has an important effect on performance, energy consumption, and guarantee of the quality of service of the computer, and serves an important role in the deployment of the virtual machine. To solve the problem of meeting the fault-tolerance requirement and guarantee high reliability of the application system based on the full use of the cloud resource under the prerequisite of various demands, the deployment framework of the feedback virtual machine in cloud platform facing the individual user’s demands of fault-tolerance level and the corresponding deployment algorithm of the virtual machine are proposed in this paper. Resource distribution strategy can deploy the virtual machine in the physical nodes where the resource is mutually complementary according to the users’ different requirements on virtual resources. The deployment framework of the virtual machine in this paper can provide a reliable computer configuration according to the specific fault-tolerance requirements of the user while considering the usage rate of the physical resources of the cloud platform. The experimental result shows that the method proposed in this paper can provide flexible and reliable select permission of faulttolerance level to the user in the virtual machine deployment process, provide a pertinent individual fault-tolerant deployment method of the virtual machine to the user, and guarantee to meet the user service in a large probability to some extent.
With the sky-rocketing development of Internet services, the power usage in data centers has been significantly increasing. This ever increasing energy consumption leads to negative environmental impact such as global warming. To reduce their carbon footprints, large Internet service operators begin to utilize green energy. Since green energy is currently more expensive than the traditional brown one, it is important for the operators to maximize the green energy usage subject to their desired long-term (e.g., a month) cost budget constraint. In this paper, we propose an online algorithm GreenBudget based on the Lyapunov optimization framework. We prove that our algorithm is able to achieve a delicate tradeoff between the green energy usage and the enforcement of the cost budget constraint, and a control parameter V is the knob to arbitrarily tune such a tradeoff. We evaluate GreenBudget utilizing real-life traces of user requests, cooling efficiency, electricity price and green energy availability. Experimental results demonstrate that under the same cost budget constraint, GreenBudget can increase the green energy usage by 11.55% compared with the state-of-the-art work, without incurring any performance violation of user requests.
Recent years have witnessed a processor development trend that integrates central processing unit (CPU) and graphic processing unit (GPU) into a single chip. The integration helps to save some host-device data copying that a discrete GPU usually requires, but also introduces deep resource sharing and possible interference between CPU and GPU. This work investigates the performance implications of independently co-running CPU and GPU programs on these platforms. First, we perform a comprehensive measurement that covers a wide variety of factors, including processor architectures, operating systems, benchmarks, timing mechanisms, inputs, and power management schemes. These measurements reveal a number of surprising observations.We analyze these observations and produce a list of novel insights, including the important roles of operating system (OS) context switching and power management in determining the program performance, and the subtle effect of CPU-GPU data copying. Finally, we confirm those insights through case studies, and point out some promising directions to mitigate anomalous performance degradation on integrated heterogeneous processors.
Reducing the power consumption has become one of the most important challenges in designing modern data centers due to the explosive growth of data. The traditional approaches employed to decrease the power consumption normally do not consider the power of IT devices and the power of cooling system simultaneously. In contrast to existing works, this paper proposes a power model which can minimize the overall power consumption of data centers by balancing the computing power and cooling power. Furthermore, an enhanced genetic algorithm (EGA) is designed to explore the solution space of the power model since the model is a liner programming problem. However, EGA is computing intensive and the performance gradually decreases with the growth of the problem size. Therefore, heuristic greedy sequence (HGS) is proposed to simplify the calculation by leveraging the nature of greed. In contrast to EGA, HGS can determine the workload allocation of a specific data center layout with only one calculation. Experimental results demonstrate that both the EGA and HGS can significantly reduce the power consumption of data centers in contrast to the random algorithm. Additionally, HGS significantly outperforms EGA in terms of the continuity of workload allocation and execution performance.
Data items are usually replicated in modern distributed data stores to obtain high performance and availability. However, the availability-consistency and latencyconsistency trade-offs exist in data replication, thus system designers intend to choose weak consistency models, such as eventual consistency, which may result in stale reads. Since stale data items may lead to serious application semantic problems, we consider how to increase the probability of data recency which provides a uniform view on recent versions of data items for all clients. In this work, we propose HARP, a framework that can enhance data recency of eventually consistent distributed data stores in an efficient and highly available way. Through detecting possible stale reads under failures or not, HARP can perform reread operations to eliminate stale results only when needed based on our analysis on write/read processes. We also present solutions on how to deal with some practical anomalies in HARP, including delayed, reordered and dropped messages and clock drift, and show how to extend HARP to multiple datacenters. Finally we implement HARP based on Cassandra, and the experiments show that HARP can effectively eliminate stale reads, with a low overhead (less than 6.9%) compared with original eventually consistent Cassandra.
In data center, applications of big data analytics pose a big challenge to massive storage systems. It is significant to achieve high availability, high performance and high scalability for PB-scale or EB-scale storage systems. Metadata server (MDS) cluster architecture is one of the most effective solutions to meet the requirements of applications in data center. Workload migration can achieve load balance and energy saving of cluster systems. In this paper, a hybrid workload migration mechanism of MDS cluster is proposed and named as HWM. In HWM, workload of MDS is classified into two categories: metadata service and state service, and they can be migrated rapidly from a source MDS to a target MDS in different ways. Firstly, in metadata service migration, all the dirty metadata of one sub file system is flushed to a shared storage pool by the source MDS, and then is loaded by the target MDS. Secondly, in state service migration, all the states of that sub file system are migrated from source MDS to target MDS through network at file granularity, and then all of the related structures of these states are reconstructed in targetMDS. Thirdly, in the process of workload migration, instead of blocking client requests, the source MDS can decide which MDS will respond to each request according to the operation type and the migration stage. The proposed mechanismis implemented in the BlueWhaleMDS cluster. The performance measurements show that the HWM mechanism is efficient to migrate the workload of a MDS cluster system and provides low-latency access to metadata and states.
Machine-learning techniques have recently been proved to be successful in various domains, especially in emerging commercial applications. As a set of machinelearning techniques, artificial neural networks (ANNs), requiring considerable amount of computation and memory, are one of the most popular algorithms and have been applied in a broad range of applications such as speech recognition, face identification, natural language processing, ect. Conventionally, as a straightforward way, conventional CPUs and GPUs are energy-inefficient due to their excessive effort for flexibility. According to the aforementioned situation, in recent years, many researchers have proposed a number of neural network accelerators to achieve high performance and low power consumption. Thus, the main purpose of this literature is to briefly review recent related works, as well as the DianNao-family accelerators. In summary, this review can serve as a reference for hardware researchers in the area of neural networks.