Supporting Isolation for Fault and Power Management with Fully Virtualized Memory Systems

Fully virtualized systems offer significant commercial advantages in certain markets by allowing users to consolidate many small servers onto a single large system that offers lower total cost and more efficient and flexible use of computing resources. With full virtualization, which means that the underlying hypervisor virtualizes processor, memory and I/O resources, the system becomes even more flexible since there is no fixed assignment of resources to partitions and operating system images. However, with such systems, the costs associated with failures and power management are much higher. For example, if a virtualized system fails, rather than losing a single system, the customer loses all of the consolidated system images, often on the order of 102 or even 103 logical servers. Similarly, unless the larger system offers power-management features comparable to those found in smaller systems, its power costs and power-related problems such as heat and potential failures are higher than the systems that it replaces. This work provides a partial solution to these problems in the area of system memory by offering a scheme for identifying and isolating if necessary the memory used by particular virtual images. It applies the well-known reverse-mapping technique to hypervisors and takes advantage of some likely characteristics of the hypervised environment to provide a compact and simple reverse mapping. With this reverse mapping, it then indicates how to reduce memory power consumption during periods of low utilization and how to isolate faulty portions of memory, affecting only those system images using the bad memory.