Disaggregation in the Cloud with μInstances and Cirrus

Resource disaggregation can provide significant improvements in the utilization of resources in the datacenter. A Google cluster trace analysis confirms that up to 70% of memory may be recovered with resource disaggregation. However, resource disaggregation in the cloud is currently unfeasible due to the hardware and network changes required by previously proposed designs. We make the observation that existing cloud offers already provide logical heterogeneity which can be leveraged to match closely the resources demanded by applications. Deploying a system across a heterogeneous set of VMs can lead to higher performance per dollar. To provide support for deploying systems in logically disaggregated clusters we propose a framework called Cirrus. Cirrus composes small compute-intensive tasks (μTasks) with large memory instances to provide the resources required by applications. Cirrus also provides an efficient interface for communication between tasks through distributed memory.