Utilisation du Cloud StratusLab : tests de performance des clusters virtuels

C. Cavet (1) , M. Le Jeune (1) , F. Dodu (1) , M. Detournay (1)

1 APC (UMR_7164) - AstroParticule et Cosmologie

Cloud computing offers IT resources on demand. This recent infrastructure in computing framework recovers three services named IaaS (Infrastructure-as-a-Service), PaaS (Platform-as-a-Service) and SaaS (Software-as-a-Service). In this work, we have focused on the IaaS layer which allows to provide virtual machine, cluster and storage on demand. We have tested the performance of the StratusLab Cloud (european research project) to determine advantages and disadvantages of Cloud computing for numerical simulation and data analysis. The academic StratusLab Cloud is a public open-source Cloud and it provides full Cloud solution: OpenNebula virtual infrastructure manager, computing resources and end-user client to instantiate and manage virtual machines. We have used classical benchmarks usually performed on supercomputers and we have compared our results with previous studies made on public paying Cloud like Amazon EC2, GoGrid and IBM. We have run benchmarks on memory bandwidth, I/O access, MPI communications, scientific applications-like and weak scalability. Furthermore, we have tested file transfer abilities with IRODS (Integrated Rule-Oriented Data System) software and we have run a numerical code to study strong scalability by measuring speedup, efficiency and Karp-Flatt metric. We have concluded that the virtual cluster of the StratusLab Cloud does not allow high performance in MPI involving tests. The slow inter-node communication is due to low virtual network i.e. virtualization layer overhead and Gigabit interconnection as shown in previous studies performed on equivalent paying Cloud. However, the non-MPI results of performed benchmarks are in agreement with classical cluster performances. Thus, the Cloud computing is an interesting solution for prototyping high performance computing applications and for running other types of scientific applications because it provides an easy and quick access to any kind of platform with reasonable resources and performances.

Domaines: Calcul parallèle, distribué et partagé [cs.DC]

C. Cavet, M. Le Jeune, F. Dodu, M. Detournay. Utilisation du Cloud StratusLab : tests de performance des clusters virtuels. Journées scientifiques mésocentres et France Grilles 2012, Oct 2012, Paris, France. ⟨hal-00766067⟩