Section: Research Program

Energy

The overall electrical consumption of DCs grows according to the demand of Utility Computing. Considering that the latter has been continuously increasing since 2008, the energy footprint of Cloud services overall is nowadays critical with about 91 billion kilowatt-hours of electricity   [91]. Besides the ecological impact, the energy consumption is a predominant criterion for providers since it determines a large part of the operational cost of their infrastructure. Among the different appraoches that have been investigated to reduce the energy footprint, some studies have been ingestigating the use of renewable energy sources to power microDCs  [64]. Workload distribution for geo-distributed DCs is also another promising approach  [66], [78], [99]. Our research will extend these results with the ultimate goal of considering the different opportunities to control the energy footprint across the whole stack (hardware and software opportunities, renewable energy, thermal management, etc.). In particular, we identified several challenges that we will address in this context within the STACK framework.

First, we propose to evaluate the energy efficiency of low-level building blocks, from the viewpoints of computation (VMs, containers, microkernel, microservices)  [58] and data (hard drives, SSD, in-memory storage, distributed file systems). For computations, in the continuity of our previous work  [56], [73], we will investigate workload placement policies according to energy (minimizing energy consumption, power capping, thermal load balancing, etc.). Regarding the data dimension, we will investigate, in particular, the trade-offs between energy consumption and data availability, durability and consistency  [51], [94]. Our ambition is to propose an adaptive energy-aware data layout and replication scheme to ensure data availability with minimal energy consumption. It is noteworthy that these new activities will also consider our previous work on DCs partially powered by renewable energy (see the SeDuCe project, in Section 6.7), with the ultimate goal of reducing the CO${}_2$ footprint.

Second, we will complete current studies to understand pros and cons of massively geo-distributed infrastructures from the energy perspective. Addressing the energy challenge is a complex task that involves considering several dimensions such as the energy consumption due to the physical resources (CPU, memory, disk, network), the performance of the applications (from the computation and data viewpoints), and the thermal dissipation caused by air conditioning in each DC. Each of these aspects can be influenced by each level of the software stack (i.e., low-level building blocks, coordination and autonomous loops, and finally application life cycle). In previous projects, we have studied and modeled the consumption of the main components, notably the network, as part of a single microDC. We plan to extend these models to deal with geo-distribution. The objective is to propose models that will enable us to refine our placement algorithms as discussed in the next paragraph. These models should be able to consider the energy consumption induced by all WAN data exchanges, including site-to-site data movements as well as the end users' communications for accessing virtualized resources.

Third, we expect to implement green-energy-aware balancing strategies, leveraging the aforementioned contributions. Although the infrastructures we envision increase complexity (because WAN aspects should also be taken into account), the geo-distribution of resources brings several opportunities from the energy viewpoint. For instance, it is possible to define several workload/data placement policies according to renewable energy availability. Moreover, a tightly-coupled software stack allows users to benefit from such a widely distributed infrastructure in a transparent way while enabling administrators to balance resources in order to benefit from green energy sources when available. An important difficulty, compared to centralized infrastructures, is related to data sharing between software instances. In particular, we will study issues raised by the distribution and replication of services across several microDCs. In this new context, many challenges must be addressed: where to place the data (Cloud, Edge) in order to mitigate dat a movements? What is the impact in terms of energy consumption, network and response time of these two approaches? How to manage the consistency of replicated data/services? All these aspects must be studied and integrated into our placement algorithms.

Fourth, we will investigate the energy footprint of the current techniques that address failure and performance variability in large-scale systems. For instance, stragglers (i.e., tasks that take a significantly longer time to finish than the normal execution time) are natural results of performance variability, they cause extra resource and energy consumption. Our goal is to understand the energy overhead of these techniques and introduce new handling techniques that take into consideration the energy efficiency of the platform  [86].

Finally, in order to answer specific energy constraints, we want to reify energy aspects at the application level and propose a metric related to the use of energy (Green SLA  [34]), for example to describe the maximum allowed CO${}_2$ emissions of a Fog/Edge service. Unlike other approaches  [67], [39], [65] that attempt to identify the best trade-off, we want to offer to developers/end-users the opportunity to select the best choice between application performance, correctness and energy footprint. Such a capability will require reifying the energy dimension at the level of big-data and interactive applications. Besides, with the emergence of renewable energy (e.g., solar panels for microDC), investigating the energy consumption vs performance trade-off  [70] and the smart usage of green energy for ICT geo-distributed services seems promising. For example, we want to offer the opportunity to developers/end-users to control the scaling of the applications based on this trade-off instead of current approaches that only considered application load. Providing such a capability will also require appropriate software abstractions.