Section: New Software and Platforms

New Software

MyNRC: image-oriented library for allocation and manipulation of SIMD 1D, 2D and 3D structures

Participant : Lionel Lacassagne.

MyNRC is multi-plateform library that can handle SSE, AVX, Neon and ST VECx registers.

CovTrack: agile realtime multi-target tracking algorithm

Participants : Michèle Gouiffès, Lionel Lacassagne, Florence Laguzet, Andrés Romero.

tmLQCD for Blue Gene/Q

Participant : Michael Kruse [correspondant] .

tmLQCD is a program suite for lattice quantum chromodynamics (Lattice QCD) using the chirally twisted Wilson quarks to reduce discretization artifacts. This software is in productive use by the European Twisted Mass Collaboration (ETMC).

As to not waste precious computation time on the supercomputers it is running on, it is important to optimize the code in order to run as fast as possible. tmLQCD has already been customized for Intel Xeon processors, the Blue Gene/L and Blue Gene/P from IBM. For the latter's successor, the Blue Gene/Q, more profound changes to the program are necessary. With these changes, tmLQCD reaches a peak performance of up to 55% of the machines theoretical floating point performance.

The Blue Gene/Q optimized tmLQCD is available at: http://github.com/Meinersbur/tmLQCD

Molly

Participant : Michael Kruse [correspondant] .

Using Polly extension, the LLVM compiler framework is able to automatically parallelize general programs for shared memory threading for by exploiting the powerful analysis and transformations of the polyhedral model.

Molly adds the ability to manage distributed memory using the polyhedral model and is therefore able to automatically parallelize even for the largest of today's supercomputer. Once the distribution of data between the computer's nodes is known, Molly determines the values that are required to be transferred between the nodes and chunks them into as few messages as possible. It also keeps tracks of the buffers required by the MPI interface. Transfers are asynchronous such that further computations take place while the data is being transferred.

Molly has not yet been officially released.

Dohko (http://dohko.io/ )

Participant : Alessandro Ferreira Leite [correspondant] .

Automating multi-cloud configuration is a difficult task. The difficulties are mostly due to clouds' heterogeneity and the lack of tools to coordinate cloud computing configurations automatically. As a result, virtual machine image (VMI) is the common approach to configure cloud environment. Although VMI can handle functional properties like minimum disk size, operating system, and software packages, it leads to a high number of configuration options, increasing the difficulty to select one that matches users' requirements. Moreover, the usage of VMI usually results in vendor lock-in. Furthermore, VMI leaves for the users the work of selecting a resource to deploy the image and for orchestrating them accordingly, i.e., the work of selecting and instantiating the VMI in each cloud. In addition, VMI migration across multiple clouds normally has a high cost due to network traffic. Finally, in case of cloud's failure, it may be difficult for users to re-create the failed environment in another cloud, since the image will be inaccessible.

Therefore, to overcome these issues, we developed a configuration management tool for cloud computing. Our tool, called Dohko, allows the users to configure a multi-cloud computing environment, following a declarative strategy. In this case, the users describe their applications and requirements and use our tool to select the resources and to set up the whole computing environment automatically, taking into account temporal and functional dependencies between the resources. Moreover, following a software product line (SPL) engineering method, Dohko captures the knowledge of configuring cloud environments in form of reusable assets. In this case, a product is a cloud computing environment that meets the user requirements, where the requirements can be either based on high or low-level descriptions. Examples of low-level descriptions include: virtualization type, disk technology, sustainable performance, among others, whereas high-level descriptions include the number of CPU cores, the RAM memory size, and the maximum monetary cost per hour. In this context, a cloud computing environment also matches cloud's configuration constraints. Besides that, thanks to the usage of an extended feature model (EFM) with attributes, our approach enables the description of the whole computing environment (i.e., hardware and software) independent of cloud provider and without requiring the usage of virtual machine image. In this case, it relies on an off-the-shelf constraint satisfaction problem (CSP) solver to implement the feature model and to select the resources.

By employing a declarative strategy, Dohko could execute a biological sequence comparison application in two distinct cloud providers (i.e., Amazon EC2 and Google Compute Engine) considering a single and a multi-cloud scenario, demanding minimal users' intervention to instantiate the whole cloud environment, as well as to execute the application. In particular, our solution tackles the lack of middleware prototypes that can support different scenarios when using services from many clouds. Moreover, it meets the functional requirements identified for multiple cloud-unaware systems  [136] such as: (a) it provides a way to describe functional and non-functional requirements through the usage of an SPL engineering method; (b) it can aggregate services from distinct clouds; (c) it provides a homogeneous interface to access services of multiple clouds; (d) it allows the service selection of the clouds; (e) it can deploy its components across many clouds; (f) it provides automatic procedures for deployments; (g) it utilizes an overlay network to connect and to organize the resources; (h) it does not impose any constraint for the connected clouds.