Sites Inria

Version française


Jean-Michel Prima - 7/04/2016

Grid'5000 available for businesses

David Margery, directeur technique de la plateforme Grid 5000

Grid'5000, the French national distributed systems and cloud computing experimental facility, has been supporting academic research since 2003. This infrastructure is now being made available to businesses interested in using it for scientific purposes, explains David Margery, who is the facility’s engineering director and a member of the Experimentation and Development department (SED) at the INRIA Rennes - Bretagne Atlantique research centre.

1,000 servers. 8,000 processing cores at 10 sites. 25 different hardware clusters. Grid'5000 is one of the French research community’s largest technical facilities. With 500 users annually and an average resource utilisation rate of 75% in 2014, this instrument is in considerable demand from the scientific community. Although primarily aimed at academia, it is now also open to enterprise.

“As we took part in European research programmes alongside industrial partners, we realised that certain companies also needed an experimentation facility ”, notes David Margery. A typical example might be an SME developing an application to send photos via mobile phone.  Before launching the product, the company would want to know how many users could send images simultaneously, and hence, how many servers the company should to deploy in order to handle the data flow. This is a textbook scaling problem. When addressing such problems relating to large data volumes, companies have a choice between a wide range of data management strategies. However, in order to determine which technological approach to adopt, they need the resources to conduct the relevant R&D correctly. Unfortunately, it is not always easy to make comparisons when using public cloud computing services such as those provided by Amazon, Google and Microsoft. “Why? ” “Because virtualisation hides the proportion of processors allocated to the user and the processing workload. Users have no detailed information about the hardware they are using. When two algorithms are tested one after the other, one might appear to be better than the other, whereas, in reality, the physical conditions may simply have been different during each processing sequence. In such cases, the analysis results are skewed. With this type of infrastructure, understanding experiment results can be a real problem. The challenge is not insurmountable, but numerous precautions must be taken. ” This is precisely the area in which Grid'5000 delivers added value.

Controlling all parameters

For the past 10 years, we have worked tirelessly to optimise the quality of experiments at the facility, in terms of their results, analysis and reproducibility. We provide bare hardware and strive to isolate parameters extraneous to the experiment. This interface between hardware deployment and experiment quality is our speciality and a field in which we have world-leading expertise. Organisations that deploy their server infrastructure and back-end on Grid'5000 know exactly what hardware they are using: which processors, which types of hard disk, etc. They even have the option to build their technology stack from scratch, in order to control every parameter. If they subsequently wish to compare two experiments, the experimental conditions will be the same in both cases. We provide comprehensive metrics, even including measuring power consumption during processing. " Taken together, these steps ensure that experiments are reproducible. Important note: “The purpose of this infrastructure is not to experimentally deploy a technology before installing it elsewhere, but to perform experiments on the technology itself. The aim is to conduct experiments with simulated users, to see what would happen in particular scenarios. ”

 From a practical perspective, access to Grid'5000 will be provided in exchange for a fee stated in core-hours. How much does access cost? “Approximately twice the price of a comparable offering on a public cloud. One of our constraints is to avoid exposing our partners to the risk of legal action on the grounds of subsidised pricing. Our approach in no way constitutes an indirect means of subsidising industrials.

Scientific support

In addition to the contract governing the use of the instrument, some companies additionally opt for a consulting contract. “In such cases, one of our scientists spends time advising the new user. ” On getting started with the equipment? “Not exactly. We have excellent tutorials for that. The consultant’s role centres more on helping the corporate user to build its experiment. Adopting the most appropriate scientific approach in order to test a hypothesis. Moving beyond a product-oriented approach that only aims to confirm that a demonstrator operates correctly. This contact with the company also offers benefits to INRIA staff, by giving them experience with real-life use cases. A scientist who discovers an unresolved problem in the course of a consulting assignment may choose to conduct further research into the issue. As we create a new interface between the research community and industry, we hope that opening up Grid'5000 will bring opportunities to develop new partnerships with our laboratories.