HPC networking is often characterized by kernel bypass, which is considered mandatory for large parallel and distributed applications. However, kernel bypass comes at a price because it breaks the traditional OS architecture, requiring applications to use special APIs and limiting the OS's control over existing network connections. We make the case that kernel bypass is not mandatory. Rather, high-performance networking relies on multiple performance-improving techniques, with kernel bypass even being detrimental to performance under specific conditions. CoRD removes kernel bypass from RDMA networks, primarily to enable efficient OS-level control over the RDMA dataplane. This control can be used to enhance security or resource allocation policies, and, as we demonstrate in one of the use cases, can improve end-to-end application performance by up to 10%. This architecture can enable Cloud-based distributed RDMA applications and facilitate deployment of coupled HPC applications.

Rising energy consumption is of growing concern for cloud data center providers. Modern processors try to counteract this problem through low-power idle states that save energy in phases with little demand for compute resources. Making proper use of this feature, however, requires knowledge about the properties of these states for the very processors used in a specific setup; most importantly, the energy consumed in each idle state and the latency for resuming normal operation. Unfortunately, hardware vendors usually do not provide this critical information.

Driven by ever increasing performance demands of compute-intensive applications, supercomputing systems comprise more and more nodes. This growth is a significant burden for fast group communication primitives and also makes those systems more susceptible to failures of individual nodes. In this paper we present a two-phase fault-tolerant scheme for group communication. Using broadcast as an example, we provide a full-spectrum discussion of our approach --- from a formal analysis to LogP-based simulations to a message-passing-based implementation running on a large cluster. Ultimately, we are able to reduce the complex problem of reliable and fault-tolerant collective group communication to a graph theoretical renumbering problem. Both, simulations and measurements, show our solution to achieve a latency reduction of 50% with up to six times fewer messages sent in comparison to existing schemes.