Section: New Results

Scaling Clouds

Efficient Docker container deployment in fog environments

Participants : Arif Ahmed, Lorenzo Civolani, Guillaume Pierre, Paulo Rodrigues de Souza Junior.

Fog computing aims to extend datacenter-based cloud platforms with additional computing, networking and storage resources located in the immediate vicinity of the end users. By bringing computation where the input data was produced and the resulting output data will be consumed, fog computing is expected to support new types of applications which either require very low network latency (e.g., augmented reality applications) or which produce large data volumes which are relevant only locally (e.g., IoT-based data analytics).

Fog computing architectures are fundamentally different from traditional clouds: to provide computing resources in the physical proximity of any end user, fog computing platforms must necessarily rely on very large numbers of small Points-of-Presence connected to each other with commodity networks whereas clouds are typically organized with a handful of extremely powerful data centers connected by dedicated ultra-high-speed networks. This geographical spread also implies that the machines used in any Point-of-Presence may not be datacenter-grade servers but much weaker commodity machines.

We investigated the challenges of efficiently deploying Docker containers in fog platforms composed of tiny single-board computers such as Raspberry PIs. Significant improvements in the Docker image cache hit rate can be obtained by sharing the caches of multiple co-located servers rather than letting them operate independently [9]. In the case when an image must be downloaded and locally installed, large performance gains can be obtained with relatively simple modifications in the way Docker imports container images [3]. Finally, we showed (in collaboration with Prof. Paolo Bellavista from the University of Bologna) that it is possible to let a container start producing useful work even before its image has been fully downloaded [14]. Another paper in this direction of work is in preparation about the way to speedup the boot phase of Docker containers. We are also exploring innovative techniques to improve the performance of live container migration in fog computing environments.

Fog computing platform design

Participants : Ali Fahs, Ayan Mondal, Nikos Parlavantzas, Guillaume Pierre, Mulugeta Tamiru.

There does not yet exist any reference platform for fog computing platforms. We therefore investigated how Kubernetes could be adapted to support the specific needs of fog computing platforms. In particular we focused on the problem of redirecting end-user traffic to a nearby instance of the application. When different users impose various load on the system, any traffic routing system must necessarily implement a tradeoff between proximity and fair load-balancing between the application instances. We demonstrated how such customizeable traffic routing policies can be integrated in Kubernetes to help transform it in a suitable platform for fog computing [15]. We extended this work to let the platform automatically choose (and maintain over time) the best locations where application replicas should be deployed. A paper on this topic is currently under submission. We finally started addressing the topic of application autoscaling such that the system can enforce performance guarantees despite traffic variations. We expect one or two publications on this topic next year.

In collaboration with Prof. Misra from IIT Kharagpur (India), and thanks to the collaboration established by the FogCity associate team, we developed mechanisms based on game theory to assign resources to competing applications in a fog computing platform. The objective of those mechanisms is to satisfy user preferences while maximizing resource utilisation. We evaluated the mechanisms using an emulated fog platform built on Kubernetes and Grid’5000, and showed that they significantly outperform baseline algorithms. A paper on this topic is in preparation.

Edgification of micro-service applications

Participants : Genc Tato, Cédric Tedeschi, Marin Bertier.

Last year, we investigated in collaboration with Etienne Riviere from UC Louvain the feasibility and possible benefits brought about by the edgification of a legacy micro-service-based application  [35]. In other words, we devised a method to classify services composing the application as edgifiable or not, based on several criteria. We applied this method to the particular case of the ShareLatex application which enables the collaborative edition of LaTeX documents. Recently, we continue this work by automate the localization and the migration of microservices. Our middleware, based on Koala  [36], a lightweight Distributed Hash Table, allows adapting compatible legacy microservices applications for hybrid core/edge deployments [21].

Community Clouds

Participants : Jean-Louis Pazat, Bruno Stevant.

Small communities of people who need to share data and applications can now buy inexpensive devices in order to use only "on premise" resources instead of public Clouds. This "self-hosting-and-sharing" solution provides a better privacy and does not need people to pay any monthly fee to a resource provider. We have implemented a prototype based on micro-services in order to be able to distribute the load of applications among devices.

However, such a distributed platform needs to rely on a very good distribution of the computing and communication load over the devices. Using an emulator of the system, we have shown that, thanks to well known optimization techniques (Particle Swarm Optimization), it is possible to quickly find a service placement resulting in a response time close to the optimal one.

This year we evaluated the results of the optimization algorithm on a prototype (5 "boxes" installed in different home locations connected by fiber or ADSL). Results shown that due to the variation of the network available bandwidth it is necessary to dynamically modify the deployment of applications. This was not a big surprise, but we were not able to find any predictive model of this variation during a day. So, we developed and experimented a dynamic adaptation of the placement of micro-services based applications based on a regular monitoring of the response time of applications. We plan to submit a paper on this topic in early 2020.

Geo-distributed data stream processing

Participants : Hamidreza Arkian, Davaadorj Battulga, Mehdi Belkhiria, Guillaume Pierre, Cédric Tedeschi.

We investigated a decentralized scaling mechanism for stream processing applications where the different operators composing the processing topology are able to take their own scaling decisions independently, based on local information. We built a simulation tool to validate the ability of our algorithm to react to load variation. Then, we started the development of a software prototype of a decentralized Stream Processing Engine including this autoscaling mechanism, and deployed it over the Grid'5000 platform. Two papers have been accepted in 2019 about this work [11], [12].

Although data stream processing platforms such as Apache Flink are widely recognized as an interesting paradigm to process IoT data in fog computing platforms, the existing performance model to capture of stream processing in geo-distributed environments are theoretical works only, and have not been validated against empirical measurements. We developed and experimentally validated such a model to represent the performance of a single stream processing operator [10]. This model is very accurate with predictions $\pm 2\%$ of the actual values even in the presence of heterogeneous network latencies. Individual operator models can be composed together and, after the initial calibration of a first operator, a reasonably accurate model for other operators can be derived from a single measurement only.

QoS-aware and energy-efficient resource management for Function-as-a-Service

Participants : Yasmina Bouizem, Christine Morin, Nikos Parlavantzas.

Recent years have seen the widespread adoption of serverless computing, and in particular, Function-as-a-Service (FaaS) systems. These systems enable users to execute arbitrary functions without managing underlying servers. However, existing FaaS frameworks provide no quality of service guarantees to FaaS users in terms of performance and availability. Moreover, they provide no support for FaaS providers to reduce energy consumption. The goal of this work is to develop an automated resource management solution for FaaS plaforms that takes into account performance, availability, and energy efficiency in a coordinated manner. This work is performed in the context of the thesis of Yasmina Bouizem. In 2019, we integrated a fault-tolerance mechanism into Fission, an open-source FaaS framework based on Kubernetes, and are currently evaluating its impact on performance, availability, and energy consumption.