MIT, An Operating System for Multicore and Clouds: Mechanisms and Implementation,” SOCC, 2010. [PDF]
Barret Rhoden, Kevin Klues, David (Yu) Zhu, Eric Brewer, “Improving Per-Node Efficiency in the Datacenter with New OS Abstractions,” SOCC, 2011. [PDF]
Factored Operating System
The Factored Operating System (FOS) proposes an OS architecture where each core runs individual microkernels that communicate using message passing. The OS has function-specific services, and each one is implemented as a parallel, distributed service. The authors argue that such an API is feasible enough to simultaneously address many-core and cloud OS requirements. Scheduling in FOS moves from time multiplexing to space multiplexing, where individual functions/programs will occupy their own cores for longer periods of time than they would have in time multiplexing.
Akaros has a more manageable goal of handling individual many-core machines instead of tackling distributed coordination. It aims for achieving high efficiency by giving bare-metal feel to applications running on top: applications have more information and greater control over resources they are using. Akaros introduces the concept of a many-core process (MCP), where all cores of an MCP are gang scheduled. MCPs are aware of cores they are running on and are never preempted without warning. Using two-level scheduling, Akaros only assign cores using space multiplexing, and MCPs are allowed to manage their own threads using time multiplexing. Akaros also provides asynchronous IO and MCPs have more information about the devices they interact with.
The need for an operating system for the datacenter is real. It is not clear which is the best OS abstraction though. To me FOS seems a long stretch, whereas Akaros is probably doable. However, neither has any reasonable performance benchmark so far.