Our first CleanSky research spotlight focused on placing virtual machines (VMs) in the cloud. In another focus area, CleanSky has investigated the larger picture as well. In many cases, customers do not rent a single VM from the service provider (that then needs to be placed appropriately), but rather require a set of VMs to run their services. One prime example are services that need to include multiple components (such as, e.g., load balancers, web servers or databases) to function properly, as shown in the figure below.
In this domain, CleanSky has focused on the management of resources in cloud environments that host multi-component services. One major trend in both industry and academia has been the introduction of edge computing, in which resources (e.g. VMs) are brought closer to the customer to facilitate lower latencies and better quality of experience for the user. This brings along a variety of follow-up questions, such as:
- Where and how should multi-component services be provisioned? Abhinandan S Prasad (ESR7) has focused on this aspect in [1,2]. He has led a team in collaboration with Nokia Bell Labs that investigated the problem from two different angles. First, an open problem was what kind of resources should be used to provision what kind of service. For instance, a database is usually more memory hungry than a webserver, which in turn needs more CPU power to handle the same request rate as the database. State-of-the-art solutions usually deploy the same type of VM for all kinds of services. In [1], Abhinandan has designed and implemented a framework that can optimally deploy complex services when there are multiple different resources available. Results show that it can increase resource utilization at the service provider by up to 50%.
Second, the introduction of edge resources raises the question of how to price these resources. For instance, running a multi-component service closer to the edge can yield better performance and an edge over competitors. Additionally, edge resources are more scarce than centralized cloud resources. Therefore, intuitively edge resources should be priced higher. The same differentiation can, for example, also be made for resources that run on renewable resources when compared with brown energy. In [2], Abhinandan tackles this issue by designing and implementing a fairly priced, time-dependent market for (edge) cloud resource allocation to multiple customers. - Deployed services will often require to process data (e.g., data analysis tasks). However, resources at the edge are limited and therefore not all data may be stored at the edge. Another critical question is thus which data should be stored closer to the edge, and which data should be stored in a central cloud? This question is addressed by our fellows Nitinder Mohan (ESR8) and Pengyuan Zhou (ESR1) together with Robert Bosch GmbH in [3], in particular for industry applications. The key idea is to design an effective caching procedure that determines, based on the workload of an application, which data should be cached at the edge. Their grouping-based cache algorithm yields significant improvements in both latency as well as cache hit ratios over non-grouping approaches.
More details can be found in the respective papers. Both research teams are currently investigating on follow-up projects based on these results.
[1] Abhinandan S Prasad (ESR7), David Koll, Jordi Arona Aroca, Jesus Omana Iglesias, Volker Hilt, Xiaoming Fu: “Optimal Resource Configuration of Complex Services in the Cloud”, to appear in Proceedings of the 17th IEEE/ACM International Symposium on Cluster, Cloud and Grid Computing. Madrid, Spain, May 14-17, 2017 [pdf][2] Abhinandan S Prasad (ESR7), Mayutan Arumaithurai, David Koll, and Xiaoming Fu: “RAERA: A Robust Auctioning Approach for Edge Resources”, to appear in Proceedings of the 1st ACM SIGCOMM Workshop on Mobile Edge Communications Networking, Los Angeles, USA, August 2017 [pdf]
[3] Nitinder Mohan (ESR8), Pengyuan Zhou (ESR1), K. Govindaraj, and Jussi Kangasharju: “Managing Data in Computational Edge Clouds”, to appear in Proceedings of the 1st ACM SIGCOMM Workshop on Mobile Edge Communications Networking, Los Angeles, USA, August 2017 [pdf]