Yank: Enabling Green Data Centers to Pull the Plug

by Ratul Mahajan

Current hypervisors offer live migration, a method to transfer the execution state (memory and disk) of a VM from one server to another, without any disruption in service. However, live migration requires the old server to stay online for a significant, undefined amount of time after migration is triggered, since a large amount of †state may need to be copied from the old to the new server. An alternative, called Remus, was proposed by Cully et al. in NSDI 2008. In Remus, an up-to-date copy of VM state is maintained on a backup server. Remus allows the old server to go offline at any instant (e.g., it can fail without warning), but its overhead in terms of network bandwidth and redundancy is high (1:1).

Yank offers an alternative between these extremes. It targets a setting in which advance warning, on the order of 5-10 seconds, is available as to when a server will go offline. The authors claim that such a warning may be available in green data centers of the future, where some "transient" servers run on unreliable, renewable energy. Once the renewable energy source goes away, a warning can be generated based on the capacity of the UPS (uninterrupted power supply). †Yank ensures that there will be no loss in state when the transient server goes offline by maintaining a slightly out-of-date version of VM state on a backup server. The amount of un-copied state on the transient server is maintained such that it can be copied to the backup server within the warning period. †In addition to bounding the amount of time within which the transient server can be taken offline, Yank needs less network bandwidth and redundancy than Remus; one backup server suffices for many (15 in the authors' experiments) transient servers.

The program committee liked the abstraction offered by Yank—managing the maximum backup time of a VM—and felt that it could be more broadly useful than its target setting. (In fact, to several reviewers, the focus on green data centers seemed a distraction as thepaper does not discuss important issues like the total power consumption of the DC and the energy implications of added redundancy.)  The paper does a good job in ironing out the details and optimizations that make Yank practical. †What remains to be seen, however, is the range of workloads for which Yank's proposed abstraction will prove most useful—these would be applications that prefer transparent durability for their state in the common case but can withstand occasional fail-stop failures (from which Yank offers no protection).