Section: New Results

Middleware for Grid and Cloud computing

Publish-Subscribe in Distributed Environments

Participants : F. Baude, F. Huet, F. Bongiovanni, L. Pellegrino, B. Sauvan, I. Alshabani, A. Bourdin, M. Antoine, A. Alshabani.

In the context of the SOA4ALL FP7-IP project, we designed and implemented a hierarchical Semantic Space infrastructure based on Structured Overlay Networks (SONS) [62] , [63] . It originally aimed at the storage and the retrieval of the semantic description of services at the Web scale [57] . This infrastructure combines the strengths of both the P2P paradigm at the architectural level and the Resource Description Framework (RDF) data model at the knowledge representation level. The achievements of this year are the following:

• In the context of the FP7 Strep PLAY and French ANR SocEDA research projects, we have been extending the aforementioned work with a content-based Publish/Subscribe abstraction in order to support asynchronous queries for RDF-based events in large scale settings, which raises some interesting challenges [26] . The goal is to build a platform for large scale distributed reasoning[25] . Such an integrated working platform [39] , [38] has been presented in two tutorials [27] , [54] .

• We have also investigated the Publish/Subscribe paradigm in the MapReduce programming model. We have proposed the concept of continuous job which allows MapReduce jobs to be re-executed when new data are added to the system. To maintain the correctness of the execution, we have introduced the notion of carried data, i.e. data which are kept between subsequent executions. An implementation has been written on top of Hadoop and a paper submitted.

Distributed algorithms for CAN-like P2P networks

Participants : L. Henrio, F. Bongiovanni, F. Huet.

The nature of some large-scale applications, such as content delivery systems or publish/subscribe systems, built on top of SONs, demands application-level dissemination primitives which do not overwhelm the overlay, i.e. efficient, and which are also reliable. Building such communication primitives in a reliable manner on top of such networks would increase the confidence regarding their behavior prior to deploying them in real settings. In order to come up with real efficient primitives, we take advantage of the underlying geometric topology of the overlay network and we also model the way peers communicate with one another. Our objective is to design and prove an efficient and reliable broadcast algorithm for CAN-like P2P networks. To this aim, in 2012 we:

• Improved the formalisation in Isabelle/HOL of a CAN-like P2P system, devised formalised tools to reason on CAN topologies, and on communication protocols on top of CANs. We designed and proved the efficiency of a first naive algorithm.

• Sketched on paper the proof of completeness and efficiency for the algorithm we designed and implemented last year.

Part of this work was done in the PhD thesis of F. Bongiovanni [10]

We are also investigating the new algorithms to efficiently build a SONs in the presence of existing data. Most of the work on SONs assume that new peers joining the network will arrive without data or fail to take into account the cost of distributing these data. Indeed, depending on the key subspace given to the new peer, some or all its data will have to be distributed in the network. In 2012:

• We proposed a first version of new join algorithms which try to allocate key sub-spaces to peers so that the amount of data that needs to be moved is minimal. An expected benefit of this work is that it should allow for fast and efficient reconstruction of a SON in case of a crash, without having to use distributed snaphshots.

Network Aware Cloud Computing

Participants : S. Malik, F. Huet.

We have worked on the Resource Aware Cloud Computing project. Its primary purpose is to address different issues which can help the scheduler to make more efficient scheduling decisions. These issues are related to the resource characteristics.

• We introduce a framework, which increases the performance of the application and ensures high level of reliability during the scheduling of application onto the cloud. It is a cloud scheduler module named as Resource Aware Cloud Scheduling (RACS) module. It helps the scheduler in making the scheduling decisions on the basis of different characteristics of cloud resources. These characteristics are reliability, network latency, and monetary cost. RACS consists of multiple sub modules, which are responsible for their corresponding tasks. In RACS, we have done the implementation for the different issues.

• We worked on a model for the reliability assessment of the cloud's computing nodes. This reliability assessment mechanism helps to do the scheduling on cloud infrastructure and perform fault tolerance on the basis of the reliability values acquired during reliability assessment. The model has different algorithms for different types of applications. Thus it has multiple reliability values for each computing node. For real time applications, the model has time based reliability assessment algorithms.

This work is part of S. Malik's PhD thesis [12]

Testbed Designs from Experimenters Requirements

Participant : F. Hermenier.

The physical design of the Emulab facility, and many other testbeds like it, has been based on the facility operators' expectations regarding user needs and behavior. If operators' assumptions are incorrect, the resulting facility can exhibit inefficient use patterns and sub-optimal resource allocation.

• We have collaborated with Robert Ricci from the University of Utah on the study of the Utah' Emulab facility to provide better testbed designs. Our study gained insight into the needs and behaviors of networking researchers by analyzing more than 500,000 topologies from 13,000 experiments submitted to Emulab.

• Using this dataset, we re-visited the assumptions that went into the physical design of the Emulab facility and considered improvements to it. Through extensive simulations with real workloads, we evaluated alternative testbeds designs for their ability to improve testbed utilization and reduce hardware costs.

The results have been published to TridentCom [22] , the reference conference related to testbeds and research infrastructures, and the article received the best paper award.

Energy Efficient Virtual Machines Placement in Data Centers

Participant : F. Hermenier.

Data centres are powerful ICT facilities which constantly evolve in size, complexity, and energy consumption. At the same time, tenants' and operators' requirements become more and more complex. The data centre operators may target different energy-related objectives while the workload volatility may alter the data centre capacity at supporting load spikes. Finally, clients of data centres are looking for dependable infrastructures that can comply with their SLA requirements.

To stay attractive, a data centre should then support these expectations. These constraints are however very specific to each of the tenants but also to the infrastructure. They also cover a large range of concerns (hardware requirements, performance, security ...) that are continuously evolving according to new trends and new technologies. Existing solutions are however ad-hoc and can not be updated easily to fit the data centres and the workload specificities.

We proposed a flexible energy-aware framework to address the multiple facets of an energy-aware consolidation of VMs in a cloud data centre.[21] This framework extended BtrPlace to make it able to address specific energy concerns. We integrated a fine grain energy model reducing either gas emissions or power consumption. We also proposed constraints to control the aggressiveness of these objectives to let the data centre reactive when a load spike occurs. We finally proposed various constraints to satisfy the hardware and the resource requirements of the tenants. The evaluation on a testbed running an industrial workload validated the practical benefits provided by the usage of our framework.

GPU-based High Performance Cloud Computing

Participants : M. Benguigui, F. Baude, F. Huet.

To address HPC, GPU devices are now considered as unavoidable cheap, energy efficient and very efficient alternative computing units. The barrier to handle such devices is the programming model: it is both very fine grained and synchronous.

Our long term goal is to devise some generic solutions in order to incorporate GPU-specific code whenever relevant into a parallel and distributed computation. The first step towards this objective is to gain some insight on how to efficiently program a non trivial but well known algorithm. We selected the American basked option pricing non embarrassingly parallel problem that was previously parallelized and distributed using ProActive master-slave approach [60] , achieving an almost linear speedup and good performances (64 CPUs based computation allowed us to solve the problem in about 8 hours). The same algorithm has been reorganized for running on a single GPU [17] and achieved the same option pricing computation in about 9 hours. The current work is to succeed to take advantage of GPUs, even if non homogeneous, hired from a Cloud or a federation of clouds at once, orchestrated by an active object acting as a GPU task delegator. The goal is to drastically lower the overall computation time for such highly time consuming stochastic simulation problems.