Virtualized Platform Management For CPU Resource Oversubscription

E-commerce services and kinds of web services are currently emerging industry directions. Compared with traditional scientific computing, the main features of these commercial services are the strong interactive and the elasticity of resources requirements. In virtualized platforms, the service providers increase the supply of resources for the virtual machines (VMs) to meet the strict performance requirements of the interactive applications. The waste of CPU is continuing as these application traffics proliferate. Eventually, the unused CPU will result in the loss of resources utilization during the service period of platform. Under new forms of applications and user experience, how to optimize the resources allocation, including the idle resource released by oversubscription strategy, is the key technology of virtualized platform management. Meanwhile, the interactivity of commercial application leads to the elasticity of resource requirement for VM. Under new situation, in order to improve the resource management, the following issues are addressed.1) how to free the idle CPU by oversubscription in the basis of the reliable performance? 2) how to detect the resource hotspot timely and accurately, and remove it by live migration in the oversubscribed platform? 3) how to implement the server consolidation strategy with performance isolation among co-hosting VMs in the oversubscribed platform?First, different applications have different characteristics and even for the same application, each of performance requirements is satisfiedby different resource need. Further, the stable application performance is the basic premise of the minimum resource allocation. To achieve this, some problems need to be solved.1) how to quantize the performance loss tolerated by users in a virtualized environment.2) how to describe the difference of performance requirements from different users? 3) how to establish a quantitative relation between performance requirements from user and CPU resource for VM? Consequently, an oversubscription scheme for CPU based on performance-aware of VM is specified. The scheme designs a method to estimate the resource oversubscription potential to guide us for resource allocation based on the corresponding performance requirements. The concrete implementation procedures of the method include quantifying the performance penalty and building a mathematical model to analyze the monitored data from VMs. This kind of resource allocation infers the lower bound of the acceptable performance level for user and then releases the idle resource to the maximum extent.Second, the performance of VM is more sensitive to the fluctuation of the workload pressure, due to the oversubscribed CPU resource. In order to meet the strict performance requirements of interactive applications, the optimization strategy of load balance based on the CPU resource oversubscription is presented. This strategy maintains the performance of VMs by two steps. The one step is monitoring the hotspots dynamically and the other step is mitigating the cascading overload caused by the improper migration. The universal way of flagging a hotspot is to construct an indicator and the corresponding threshold for each VM. Taking the workload pressure and the performance requirement into consideration, the value of this threshold is varied for the purpose of avoiding aggressive or conservative migrations. At the same time, for the short life cycle application, the hotspots, detected at the end of the program execution, will be optimized. The hotspot is responsible for signaling a need for VM migration when the resource competition or performance violation is detected. The cascading overload may be triggered when the VM migration is performed haphazardly and further resource shortage will be occurred on the destination physical machines (PMs). This kind of overload is mitigated by controlling the migration and estimating the resources for destination PM candidates accurately.Finally, the fragmentation of physical resource is caused by the scattered distribution of oversubscribed VMs. Therefore, the server consolidation is employed and it is a pity that there are some problems in the oversubscribed environment:1) Increasing oversubscribed VMs host in the same PM and then the performance isolations among co-hosting VMs are destroyed by the underlying resources starvation.2) The migration cost is high because the server consolidation strategy is triggered more frequently.3) The oversubscription strategy is failure for some VMs. To address these problems above, the optimization strategy of server consolidation based on the CPU resource oversubscription is presented. In this strategy, the performance isolations are protected by controlling the consolidation density for PMs and the online first fit algorithm is used to optimize the implementation of consolidation strategy. The VMs, which have the same performance requirements, are placed in the same PM to prevent the failure of oversubscription.In summary, a series of resources management strategies are proposed, including resource allocation by oversubscription, load balance, and VM placement. They cooperate with each other in order to promote the service capability in the virtualized platform.