Enforcing CPU allocation in a heterogeneous IaaS

• We identify issues related to changing underlying hardware in public IaaS.
• We propose a solution for DVFS and VM migration across heterogeneous nodes.
• We provide an implementation of our solution within the Xen hypervisor.
• The evaluation of our prototype confirms the accuracy of our solution.

In an Infrastructure as a Service (IaaS), the amount of resources allocated to a virtual machine (VM) at creation time may be expressed with relative values (relative to the hardware, i.e., a fraction of the capacity of a device) or absolute values (i.e., a performance metric which is independent from the capacity of the hardware). Surprisingly, disk or network resource allocations are expressed with absolute values (bandwidth), but CPU resource allocations are expressed with relative values (a percentage of a processor). The major problem with CPU relative value allocations is that it depends on the capacity of the CPU, which may vary due to different factors (server heterogeneity in a cluster, Dynamic Voltage Frequency Scaling (DVFS)). In this paper, we analyze the side effects and drawbacks of relative allocations. We claim that CPU allocation should be expressed with absolute values. We propose such a CPU resource management system and we demonstrate and evaluate its benefits.