| Infrastructure-as-a-Service (IaaS) cloud computing provides customers with com-puting resources in the form of virtual machines (VMs) and scales the leased resources on demand, while charging the customers by a simple "pay-as-you-go" pricing model. Ac-cordingly, such an emerging computing paradigm can save the budget of workload execu-tion for customers, and improve the utilization of datacenters for cloud providers. Due to the excellent characteristics above, IaaS cloud computing has gained much popularity in both the academy and industry. Currently, a number of big IT enterprises, such as Amazon, Google, and Microsoft, have released their respective cloud computing solutions, sequen-tially, in order to allow more individuals and companies to fast deploy, maintain or scale up their business services while cutting down the monetary cost of their business opera-tions.However, the VM instances in IaaS clouds need to share the physical resources in cloud datacenters with other VMs, including hardware resources of physical machines (PMs), such as CPU, cache and I/O resources, as well as the shared network and storage resources in datacenters. As a result, there exists severe VM contention on the shared computing resources, thereby causing significant performance degradation and variation, i.e., performance unpredictability, to the VM workloads running in IaaS clouds. Such a performance issue further hinders the customers from moving their performance-sensitive business services to the cloud, and consequently restricts the application scope of the IaaS cloud, which has become one of the main obstacles to the development of cloud comput-mg.Although there have been a number of preliminary works on optimizing and guaran-teeing the VM performance in IaaS clouds, a series of critical challenges remain to be ad-dressed. First, in the backend of IaaS clouds, there is a lack of a holistic model to compre-hensively quantify the performance interference on the multi-dimensional resources of VMs, which can be further leveraged to make the performance-optimal choice for VM migrations. Second, in the frontend of IaaS clouds, existing VM provisioning strategies are oblivious to the performance heterogeneity of VMs, which is caused by the hardware heterogeneity of VM instances and VM interference in the cloud. Third, in the workload execution environment of IaaS clouds, existing task scheduling mechanisms of MapRe-duce are oblivious to the heterogeneity of network available bandwidth of racks in shared clusters, e.g., IaaS clouds, which is very likely to cause network hotspots issues in data-centers. To address the challenges above, we propose a series of performance models and optimization techniques to solve the performance issues of VM workloads in the three cloud levels (i.e., the backend, frontend, and workload execution environment of IaaS clouds), for achieving predictable performance in the IaaS cloud.First, we propose an interference-aware VM migration strategy, named as iAware, to mitigate the VM interference in IaaS cloud backend. Through realistic experiments with mixed types of representative workloads in IaaS clouds, we identify and capture a series of key system-level factors that affect the performance interference of VMs, including mul-ti-dimensional resource utilization of VMs and PMs. Using the key factors above as well as the principle of demand and supply in microeconomics, we design a lightweight mul-ti-resource demand-supply performance model, in order to quantify the migration inter-ference and co-location interference of VMs online. Based on such a model, the iAware strategy can comprehensively optimize the two kinds of VM interference in a holistic manner, and make the migration decision with the minimum performance interference of VMs. Moreover, the iAware strategy can cooperate with existing VM migration or con-solidation algorithms, in order to achieve the goals of VM migrations, e.g., load balancing and power saving, while minimizing the performance interference of VMs.Second, we propose a network performance-aware task scheduling mechanism of MapReduce, name as Net-Aware, to solve the network hotspots issue in the execution en-vironment of workloads (i.e., shared clusters in IaaS clouds). Through theoretically mod-eling the relationship between the completion time of a MapReduce job and the task as-signment in racks, we analyze the performance penalty and bonus which are generated by the scheduling of map and reduce tasks. Using the model and analysis above, we design two greedy heuristics, which optimize the assignment of map tasks and reduce tasks, re-spectively, in the racks of shared clusters. Such task scheduling can alleviate the network hotspots caused by the heterogeneous network bandwidth available in racks and the shuf-fling of large amounts of data in shared clusters. Net-Aware task scheduling mechanism integrates the two proposed algorithms to improve the performance of MapReduce jobs and minimize the job completion time.Finally, based on the IaaS cloud platform optimized by iAware and Net-Aware, we propose a performance heterogeneity-aware VM provisioning strategy, named as Heifer, to further alleviate the performance variation of VMs leased in the frontend of IaaS clouds. Through realistic experiments with MapReduce which is the representative workloads in IaaS clouds, we analyze the two main causes of the performance variation of VM work-loads, i.e., the hardware heterogeneity and interference of VMs. Furthermore, based on the CPU and I/O scheduling mechanisms in Xen, we analyze the relationship between the re-source utilization of VMs and the performance of VM workloads. Using the online meas-ured resource utilization of VMs and VM interference model in iAware, we devise a per-formance prediction model of MapReduce applications, by explicitly considering the hardware heterogeneity and performance interference of VM instances. Based on such a performance model, Heifer provisioning strategy can select the hardware type of VM in-stances with the optimal running performance of workloads and lease the least number of VMs for customers.In summary, we develop a series of performance optimization and guarantee tech-niques for VM workloads, to solve the performance issues in different levels of IaaS clouds. The proposed techniques are demonstrated to be effective in alleviating the per-formance degradation and variation of VMs, thereby providing predictable performance for VM workloads in IaaS clouds. In addition, these techniques can save the job budget for customers and improve the job throughput and revenue of datacenters for cloud providers. |
PDF Full Text Request