Towards Superclouds

ion Level Feature Existing Clouds Application monitoring Auto-scaling Non-live migration Mutable Hypervisor Page sharing [81, 145] Overdriver [154] Revirt [65] Remus [61] Live migration [57] Cross-provider live migration Exposed Hardware vSnoop [92] Superpages [121] Page coloring [93] Non fate-sharing Unsupported paravirtualization Table 3.1: Cloud abstractions and extensions they enable maximally utilize their VMs through techniques like page sharing [81, 145] or aggressive oversubscription [154]. Live VM migration [57] between multiple clouds—public or private—is virtually impossible. Innovative hypervisor-level techniques for high availability [61] or intrusion detection [65] are unavailable, while further customization and experimentation at this level is stifled. Buried Hardware: Existing clouds bury the details of hardware beneath a virtual machine abstraction. Users must depend on the provider to expose everything from efficient I/O interfaces to physical fate-sharing information. Moreover, users cannot implement hardware-dependent tricks to squeeze the best performance out of the rented resources. Time-sensitive tasks, such as TCP acknowledgment [92], are difficult. Superpage [121] utilization is not efficient on virtual memory that may not be contiguous, and performance opportunities like page coloring [93] are also lost.