Section: New Results

Energy-aware computing

Renewable energy

In his PhD thesis [1], Md Sabbir Hasan proposes – across three different contributions – how to smartly use green energy at the infrastructure and application levels for further reduction of the corresponding carbon footprints. First, he investigates the options and challenges to integrate different renewable energy sources in a realistic way and proposes a Cloud energy broker, which can adjust the availability and price combination to buy Green energy dynamically from the energy market in advance to make a data center partially green. Then, he introduces the concept of virtualization of green energy, which can be seen as an alternative to energy storage used in data centers to eliminate the intermittency problem to some extent. With the adoption of this virtualization concept, we can maximize the usage of green energy contrary to energy storage which induces energy losses, while introducing a notion of Green Service Level Agreement based on green energy for service provider and end-users. Finally, he proposes an energy adaptive autoscaling solution to exploit application internals to create green energy awareness in the interactive SaaS applications, while respecting traditional QoS properties.

In [9], we present a scheme for green energy management in the presence of explicit and implicit integration of renewable energy in data center. More specifically we propose three contributions: i) we introduce the concept of virtualization of green energy to address the uncertainty of green energy availability, ii) we extend the Cloud Service Level Agreement (CSLA) language (http://web.imt-atlantique.fr/x-info/csla) to support Green SLA by introducing two new threshold parameters and iii) we introduce green SLA algorithm which leverages the concept of virtualization of green energy to provide per interval specific Green SLA. Experiments were conducted with real workload profile from PlanetLab and server power model from SPECpower to demonstrate that Green SLA can be successfully established and satisfied without incurring higher cost.

In [8], we investigate a thorough analysis of energy consumption and performance trade-off by allowing smart usage of green energy for interactive cloud application. Moreover, we propose an auto-scaler, named as SaaScaler, that implements several control loop based application controllers to satisfy different performance (i.e., response time, availability and user experience) and resource aware metrics (i.e., quality of energy). Based on extensive experiments with RUBiS benchmark and real workload traces using single compute node in Openstack/Grid'5000, results suggest that 13% brown energy consumption can be reduced without deprovisioning any physical or virtual resources at IaaS layer while 29% more users can access the application by dynamically adjusting capacity requirements. In [23], we add to the previous paper the capability of the infrastructure layer to be elastic. We propose a PaaS solution which efficiently utilize the elasticity nature at both infrastructure and application levels, by leveraging adaptation in facing to changing condition i.e., workload burst, performance degradation, quality of energy, etc. While applications are adapted by dynamically re-configuring their service level based on performance and/or green energy availability, the infrastructure takes care of addition/removal of resources based on application's resource demand. Both adaptive behaviors are implemented in separated modules and are coordinated in a sequential manner. We validate our approach by extensive experiments and results obtained over Grid'5000 testbed. Results show that, application can reduce significant amount of brown energy consumption by 35% and daily instance hour cost by 37% compared to a baseline approach.

in [28] we address the problem of improving the utilization of renewable energy for a single data center by using two approaches: opportunistic scheduling and energy storage. Our first result deals with analyzing the workload to find ideal solar panel dimension and battery size, this is used to power the entire workload without any brown energy consumption. However, in reality, either the solar panel dimension or the battery size are limited, and we still have to address the problem of matching the workload consumption and renewable energy production. The second result shows that opportunistic scheduling can reduce the demand for battery size while the renewable energy is sufficient. The last results demonstrate that for different battery sizes and solar panel dimensions, we can find an optimal solution combining both approaches that balances the energy losses due to different causes such as battery efficiency and VM migrations due to consolidation algorithms.

In [5] we presented the EPOC project, focus on energy-aware task execution from the hardware to application’s components in the context of a mono-site data center (all resources are in the same physical location) which is connected to the regular electric Grid and to renewable energy sources (such as windmills or solar cells). we have presented the EpoCloud principles, architecture and middleware components. EpoCloud is our prototype, which tackles three major challenges: 1) To optimize the energy consumption of distributed infrastructures and service compositions in the presence of ever more dynamic service applications and ever more stringent availability requirements for services; 2) To design a clever cloud’s resource management, which takes advantage of renewable energy availability to perform opportunistic tasks, then exploring the trade-off between energy saving and performance aspects in large-scale distributed system; 3) To investigate energy-aware optical ultra high-speed interconnection networks to exchange large volumes of data (VM memory and storage) over very short periods of time.

in [31] we extend our previous work on PIKA (focus 2 in the EPOC project) and introduced the green energy aware scheduling problem (GEASP) to optimize the energy consumption of a small/medium size data center. Using our model to solve the GEASP, we could optimize the energy consumption of a small/medium size data center in three ways. First, we slightly decrease its overall energy consumption, second we considerably decrease its brown energy consumption and finally we significantly increase its green energy consumption.

Energy-aware consolidation and reconfiguration

In [41] we compared the performance of VMs and containers when consolidating multiple services, in terms of QoS and EE. Our experiments compared two broadly recognized virtualization technologies: KVM for the VM approach, and Docker for the containers. We conclude that Docker outperforms KVM both in QoS and EE. According to our measurements, Docker allows running up to a 21% more services than KVM, when setting a maximum latency of 3,000 ms. In this configuration, Docker offers this service while using a 11.33% less energy than KVM. At a datacenter level, the same computation could run using less servers and less energy per server, accounting for a total of a 28% energy savings inside the datacenter.

The emergence of Internet of Things (IoT) is participating to the increase of data- and energy-hungry applications. As connected devices do not yet offer enough capabilities for sustaining these applications, users perform computation offloading to the cloud. To avoid network bottlenecks and reduce the costs associated to data movement, edge cloud solutions have started being deployed, thus improving the Quality of Service. In [29] , we advocated for leveraging on-site renewable energy production in the different edge cloud nodes to green IoT systems while offering improved QoS compared to core cloud solutions. We proposed an analytic model to decide whether to offload computation from the objects to the edge or to the core Cloud, depending on the renewable energy availability and the desired application QoS. This model is validated on our application use-case that deals with video stream analysis from vehicle cameras.

In [33], we address the problem of stragglers (i.e., slow tasks) in Big Data applications. In particular, we introduce a novel straggler detection mechanism to improve the energy efficiency of speculative execution in Hadoop, namely a hierarchical detection mechanism. The goal of this detection mechanism is to identify critical stragglers which strongly affect the job execution times and reduce the number of killed speculative copies which lead to energy waste. We also present an energy-aware copy allocation method to reduce the energy consumption of speculative execution. The core of this allocation method is a performance model and an energy model which expose the trade-off between performance and energy consumption when scheduling a copy. We evaluate our hierarchical detection mechanism and energy-aware copy allocation method on the Grid'5000 testbed using three representative MapReduce applications. Experimental results show a good reduction in the resource wasted on killed speculative copies and an improvement in the energy efficiency compared to state-of-the-art mechanisms.

The increasing size of main memories has lead to the advent of new types of storage systems. These systems propose to keep all data in distributed main memories. In [35], we present a study to characterize the performance and energy consumption of a representative in-memory storage system, namely RAMCloud, to reveal the main factors contributing to performance degradation and energy-inefficiency. Firstly, we reveal that although RAMCloud scales linearly in throughput for read-only applications, it has a non-proportional power consumption. Mainly because it exhibits the same CPU usage under different levels of access. Secondly, we show that prevalent Web workloads i.e., read-heavy and update-heavy workloads, can impact significantly the performance and the energy consumption. We relate it to the impact of concurrency, i.e., RAMCloud poorly handles its threads under highly-concurrent accesses. Thirdly, we show that replication can be a major bottleneck for performance and energy. Finally, we quantify the overhead of the crash-recovery mechanism in RAMCloud on both energy-consumption and performance.