HAL is a multidisciplinary open access archive for the deposit and dissemination of scientific research documents, whether they are published or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers. L'archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d'enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.

Delivering on-demand web content to end-users in order to carry out strict QoS metrics is not a trivial task for globally distributed network providers. This task becomes still harder when content popularity varies over the time and the SLA definitions have to include both transfer rate and latency metrics. Current worldwide content delivery approaches and datacenter infrastructures rely on cumbersome replication schemes that are agnostic to edge-network resources, and damage content provision. In this work we present AREN, an novel replication scheme for cloud storage on edge networks. AREN relies on a collaborative cache strategy and bandwidth reservation to adapt the replication degree according to strict SLA contracts and content popularity growth. We have evaluated the performances of replication schemes on edge networks using Caju, a content distribution system for edge networks. Compared to a noncollaborative caching, evaluations show that AREN prevents nearly 99.8% of all SLA violations when the storage system is heavily loaded. We also show that AREN provides a sevenfold decrease in the amount of storage usage for replicas, and it increases by roughly 20% the aggregate bandwidth, hence accelerating content delivery.

Scale-out storage systems (SoSS) have become increasingly important for meeting availability requirements of web services in cloud platforms. To enhance data availability, SoSS rely on a variety of built-in fault-tolerant mechanisms, including replication, redundant network topologies, advanced request scheduling, and other failover techniques. However, performance issues in cloud services still remain one of the main causes of discontentment among their tenants. In this paper, we propose an anomaly detection approach for SoSS that predicts cloud anomalies caused by memory and network faults. To evaluate our prediction model, we built a testbed simulating a virtual data center using VMware. Experimental results confirm that the injected faults are likely to undermine the data availability in SoSS. They suggest that although unsupervised learning has been the most common method for anomaly detection, a supervisedbased implementation of the same model reduces the false positive rate by roughly 10%. Our analysis also points out that probing SoSS-specific monitoring data at the VM-level contributes to improve the anomaly prediction efficiency.

No abstract