llamaOS: A Solution for Virtualized High-Performance Computing Clusters

For most use cases, the benefits of modern virtualization technologies vastly exceed the impact of the runtime computational overhead they introduce. However, for high-performance computing (HPC) systems, the additional overhead is often detrimental to critical elements of application performance. This work presents llamaOS (Low-LAtency Minimal Appliance Operating System) as a new approach to circumvent the overhead of virtualization and provide performance capabilities superior to even equivalent native (nonvirtual) systems. Within the llamaOS virtual environment, traditional system level services can be optimized for a specific target application domain. For example, a specialized networking framework, named llamaNET, is designed specifically for low-latency message passing communication. Performance studies show llamaNET provides a 70% reduction in network latency over native TCP for message sizes less than 8 kilobytes, while at the same time providing slightly higher communication throughput for larger sizes. This improvement in communication performance directly translates to overall application speedup for software that exhibits fine-grain parallelism.