51±¾É«

Dependable operation of a power system requires fast-acting system reserves provided by flexibility resources that can follow a reference signal in an accurate, timely, and dependable manner. In energy cells, an incom- ing reference signal is received by the control unit of the energy cell. The reference signal is then disaggregated to setpoints for individual flexibility resources. These individual setpoints are sent via a communication network to the flexibility resources, which realize the requested change in power. Generally, the components on the path from the control unit to the flexibility resources may fail. That is to say, a failure may hinder or halt the operation of the control unit, the communication network, or the individual flexibility resources. From the viewpoint of an energy cell, component failures are faults that should be tolerated. This paper proposes an approach for accurate, timely, and dependable control of an energy cell despite faults. The proposed approach uses a genetic algorithm to estimate the state of failed flexibil- ity resources, combined with either optimal or proportional activation of flexibility resources, and state-machine replication. The results show the performance of the approach on a benchmark grid using realistic reference signals and realistic performance requirements based on the example of automatic frequency restoration reserves. The optimal activation of assets uses less energy to follow the reference signal, whereas the proportional activation of assets reduces the variance in the impact depending on which flexibility resources fail. The state estimation explicitly treats the different ways assets can fail and outperforms a comparable approach based only on estimating fluctuations of power provided by flexibility resources. The state-machine replication is shown to withstand crashes of replicas of the control unit with only small delays in reaction.

State-machine replication is an established concept to build fault-tolerant services, whereby a consensus protocol is used to enforce a deterministic request order throughout a set of redundant replicas. This ensures that state-altering requests are executed in the same order throughout all replicas. There is an established read-only optimization, which allows read requests to bypass the consensus protocol to reduce the end-to-end latency, but requires the clients to wait for more responses for all requests, including ordered requests, to guarantee linearizability. In this paper, we propose a novel approach that introduces speculatively-ordered read (SOR) requests and allows to reduce the required response quorum for ordered requests, while still preserving linearizability. We conducted a series of experiments with wide-area and cluster deployments, which show that our approach can significantly reduce the number of read-write conflicts and thereby drastically improve the request processing latency of ordered requests.

State-machine replication (SMR) is a popular fault-tolerance technique for building highly-available services. Usually, consensus protocols are used to enforce a deterministic service-request ordering among replicas, in order to prevent their state from diverging. Over the last decades, a multitude of consensus protocols have been developed which come with different characteristics but also with different communication and programming models. Our Consensus-Agnostic Replication Toolkit (CART) is a wrapper for consensus protocols that relieves clients from most consensus configuration and support. Besides, it implements a generic client and application interface to support different consensus protocols and configurations, e.g. in cloud deployments. CART has built-in authentication of services based on BLS threshold signatures. It can further prove malicious behaviour of replicas, thus speeding up recovery in case of Byzantine faults. We evaluate the performance overhead of our approach in a real-world WAN deployment for two different consensus protocol implementations using the YCSB benchmark. Our results show that CART is able to reach up to 90% of the throughput achieved by the native consensus protocol with an additional latency overhead of only 10%.