Section: New Results

Structuring of Applications for Scalability

Participants : Sylvain Contassot-Vivier, Thomas Jost, Jens Gustedt, Soumeya Leila Hernane, Constantinos Makassikis, Stéphane Vialle.

Large Scale and Interactive Fine Grained Simulations

Our library parXXL allows the validation of a wide range of fine grained applications and problems. We were able to test the interactive simulation of PDEs in physics, see [5] , on a large scale. Also, biologically inspired neural networks have been investigated using parXXL and the InterCell software suite. The InterCell suite and these applicative results have been presented in [29] .

Large Scale Models and Algorithms for Random Structures

A realistic generation of graphs is crucial as an input for testing large scale algorithms, theoretical graph algorithms as well as network algorithms, e.g platform generators.

Commonly used techniques for the random generation of graphs have two disadvantages, namely their lack of bias with respect to history of the evolution of the graph, and their incapability to produce families of graphs with non-vanishing prescribed clustering coefficient. In this work we propose a model for the genesis of graphs that tackles these two issues. When translated into random generation procedures it generalizes well-known procedures such as those of Erdős & Rény and Barabási & Albert. When just seen as composition schemes for graphs they generalize the perfect elimination schemes of chordal graphs. The model iteratively adds so-called contexts that introduce an explicit dependency to the previous evolution of the graph. Thereby they reflect a historical bias during this evolution that goes beyond the simple degree constraint of preference edge attachment. Fixing certain simple statical quantities during the genesis leads to families of random graphs with a clustering coefficient that can be bounded away from zero.

A journal article describing intensive simulations of these models that confirm the theoretical results and that show the ability of that approach to model the properties of graphs from application domains has been published as [13] .

Development environment for co-processing units

In the framework of the PhD thesis of Wilfried Kirschenmann, co-supervised by Stéphane Vialle (SUPELEC & AlGorille team) and Laurent Plagne (EDF SINETICS team), we have designed and implemented a unified framework based on generic programming to achieve a development environment adapted both to multi-core CPUs, multi-core CPUS with SSE units, and GPUs, for linear algebra applied to neutronic computations. Our framework is composed of two layers: (1) MTPS is a low-level layer hiding the real parallel architecture used, and (2) Legolas++ is a high-level layer allowing the application developer to rapidly implement linear algebra operations. The Legolas++ layer aims at decreasing the development time, while the MTPS layer aims at automatically generating very optimized code for the target architecture, thus leading to decreased execution times. Experimental performances of the MTPS layer appeared very good, the same source code achieved performances close to 100% of the theoretical ones, on any of the supported target architectures. Our strategy is to generate optimized data storage and data access code for each target architecture, not just different computing codes.

A new version of Legolas++ is under development, and a minimal version has been implemented in 2011. It is optimized to use the MTPS layer: source code is generic while an optimized code is automatically generated to efficiently use all SSE/AVX vector units of a multicore CPU. An article on that work is accepted in the post-proceedings of PARA 2010 and will be published at the end of 2011; the thesis of Wilfried Kirschenmann will be defended in early 2012.

Structuring algorithms for co-processing units

Since 2009, we have designed and experimented several algorithms and applications, in the fields of option pricing for financial computations, generic relaxation methods, and PDE solving applied to a 3D transport model simulating chemical species in shallow waters. We aim at designing a large range of algorithms for GPU cluster architectures, to develop a real knowledge about mixed coarse and fine grained parallel algorithms, and to accumulate practical experience about heterogeneous cluster programming.

Our PDE solver on GPU cluster has been designed in the context of a larger project on the study of asynchronism (see  3.1 and  6.1.5 ). The iterations of the asynchronous parallel algorithm runs faster, but it requires more iterations and a more complex detection of convergence, see Section  6.1.5 below. We measured both computing and energy performances of our PDE solver in order to track the best solution, in function of the problem size, the cluster size and the features of the cluster nodes. We are tracking the most efficient solution for each configuration. It can be based on a CPU or a GPU computing kernel, and on a synchronous or asynchronous parallel algorithm. Moreover, the fastest solution is not always the less energy consuming. Our recent results are introduced in [26] , and in an article accepted in the post-proceedings of PARA 2010. In 2011 we improved our asynchronous implementation. However, the most asynchronous version has led to significantly more complex code (with an increased probability of remaining bugs) but to similar performances. At the opposite, we designed and implemented different convergence detection mechanisms in our asynchronous version, and some versions seem to achieve really better performances. Execution time and energy consumption performances have now to be measured again for many configurations. We aim to get new complete performance evaluation at the beginning of 2012. Then we will design an automatic selection of the right kernel and the right algorithm, and we will implement an auto-setting application function of a global instruction of the user (to achieve a fast run, or a low consumption run, or a compromise...).

At last, we have continued to design option pricers on clusters of GPUs, with Lokman Abbas-Turki (PhD student at University of Marne-la-Valée) and some colleagues from financial computing. In the past we developed some European option pricers, distributing independent Monte-Carlo computations on the nodes of a GPU cluster. In 2010 we succeeded to develop an American Option pricer on our GPU clusters, distributing strongly coupled Monte-Carlo computations. The Monte-Carlo trajectories depend on each others, and lead to many data transfers between CPUs and GPUs, and to many communications between cluster nodes. First results were encouraging, we achieve speedup and size up. In 2011 we optimized a major step of our algorithm, consisting in a 4D to 2D reduction on GPU. Performances have increased, and are significantly easier to achieve. The configuration tuning of the application, function of the problem size and the number of computing nodes, has been simplified. Again, we investigate both computing and energy performances of our developments, in order to compare interests of CPU clusters and GPU clusters considering execution speed and the exploitation cost of our solution.


In the previous paragraph is mentioned a project including the study of sparse linear solvers on GPU. That project deals with the study of asynchronism in hierarchical and hybrid clusters mentioned in  3.1 .

In that context, we study the adaptation of asynchronous iterative algorithms on a cluster of GPUs for solving PDE problems. In our solver, the space is discretized by finite differences and all the derivatives are approximated by Euler equations. The inner computations of our PDE solver consist in solving linear equations (generally sparse). Thus, a linear solver is included in our solver. As this part is the most time consuming, to decrease the overall computation time it is essential to get a version that is as fast as possible. This is why we have decided to implement it on GPU, as discussed in the previous paragraph. Our parallel scheme uses the Multisplitting-Newton which is a more flexible kind of block decomposition. In particular, it allows for asynchronous iterations.

Our first experiments, conducted on an advection-diffusion problem, have shown very interesting results in terms of performances [8] . However, we investigate the possibility to insert periodic synchronous iterations inside the asynchronous scheme in order to improve the convergence detection delay. This is especially interesting on small/middle clusters with efficient networks.

Moreover, another aspect which is worth being studied is the full use of all the computational power present on each node, in particular the multiple cores, in conjunction with the GPU. This is still a work in progress.

New Control and Data Structures for Efficiently Overlapping Computations, Communications and I/O

With the thesis of Pierre-Nicolas Clauss we introduced the framework of ordered read-write locks, ORWL, see [3] . These are characterized by two main features: a strict FIFO policy for access and the attribution of access to lock-handles instead of processes or threads. These two properties allow applications to have a controlled pro-active access to resources and thereby to achieve a high degree of asynchronism between different tasks of the same application. For the case of iterative computations with many parallel tasks which access their resources in a cyclic pattern we provide a generic technique to implement them by means of ORWL. It was shown that the possible execution patterns for such a system correspond to a combinatorial lattice structure and that this lattice is finite iff the configuration contains a potential deadlock. In addition, we provide efficient algorithms: one that allows for a deadlock-free initialization of such a system and another one for the detection of deadlocks in an already initialized system.

We have developed a standalone distributed implementation of the API that is uniquely based on C and POSIX socket communications. Our goal is to simplify the usage of ORWL and to allow portability to a large variety of platforms. This implementation runs on different flavors of Linux and BSD, on different processor types Intel and ARM, and different compilers, gcc, clang, opencc and icc. An experimental evaluation of the performance is on its way. An engineering support from the local INRIA center has allowed to advance this implementation and to perform intensive benchmarks. The results have been presented in [28] .

Data Handover, DHO, is a general purpose API that combines locking and mapping of data in a single interface. The access strategies are similar to ORWL, but locks and maps can also be hold only partially for a consecutive range of the data object. It is designed to ease the access to data for client code, by ensuring data consistency and efficiency at the same time.

In the thesis of Soumeya Hernane, we use the Grid Reality And Simulation (GRAS) environment of SimGrid, see  5.4 , as a support for an implementation of DHO. GRAS has the advantage of allowing the execution in either the simulator or on a real platform. A first series of tests and benchmarks of that implementation demonstrates the ability of DHO to provide a robust and scalable framework, [18] . A step forward towards a distributed algorithm that allows distributed read-write locks with dynamic participation of process has been achieved in [30] .

Energy performance measurement and optimization

Several experiments have been done on the GPU clusters of SUPÉLEC with different kinds of problems ranging from an embarrassingly parallel one to a strongly coupled one, via some intermediate levels. Our first results tend to confirm our first intuition that the GPUs are a good alternative to CPUs for problems which can be formulated in a SIMD or massively multi-threading way. However, when considering not embarrassingly parallel applications the supremacy of a GPU cluster tends to decrease when the number of nodes increases. This observation was the starting point of our participation to the COST-IC0804 about energy efficiency in large scale distributed systems, and an article accepted in the post-proceedings of PARA 2010 introduces our results achieved with our PDE solver distributed on our GPU clusters.

In 2011 we conducted new experiments and optimizations of our PDE solver and our American option pricer on the GPU clusters of SUPÉLEC. These experiments are still ongoing, and the optimization of this software should be achieved at the beginning of 2012. Simultaneously, we designed the foundations of a complete software architecture of self-configuring applications, choosing the right compute kernel and the right parallel algorithm to use, automatically. The global objectives to respect would either the overall speed, low energy consumption, or a speed-energy compromise. This global objective can be set by the user, or by an intelligent scheduler that aims to optimize a set of runs on a large cluster. This software architecture foundation has been introduced to a COST-IC0804 meeting in Budapest in June 2011.

In order to achieve this goal we need to establish some models of energy consumption of our applications on our CPU+GPU clusters, to be able to implement some heuristic of auto-setting. In 2011 we published a book chapter [26] introducing our first modeling strategies. Next step will be to implement a first auto-setting application, and to experiment its performances.

Load balancing

A load-balancing algorithm based on asynchronous diffusion with bounded delays has been designed to work on dynamical networks  [34] . It is by nature iterative and we have provided a proof of its convergence in the context of load conservation. Also, we have given some constraints on the load migration ratios on the nodes in order to ensure the convergence. This work has been extended, especially with a detailed study of the imbalance of the system during the execution of a parallel algorithm simulated in the SimGrid platform.

The perspectives of that work are double. The first one concerns the internal functioning of our algorithm. There is an intrinsic parameter which tunes the load migration ratios and we would like to determine the optimal value of that ratio. The other aspect is on the application side in a real parallel environment. Indeed, we are currently applying this algorithm to a parallel version of the AdaBoost learning algorithm. This will allow us to study the best parameter to choose and to compare our load-balancing scheme to other existing ones.

Concerning the Neurad project, our parallel learning proceeds by decomposing the data-set to be learned. However, using a simple regular decomposition is not sufficient as the obtained sub-domains may have very different learning times. Thus, we have designed a first domain decomposition of the data set yielding sub-sets of similar learning times  [40] . One of the main issue in this work has been the determination of the best estimator of the learning time of a sub-domain. As the learning time of a data set is directly linked to the complexity of the signal, several estimators taking into account that complexity have been tested, among which the entropy. However, the entropy is not the best estimator in that context, and we had to design a specific estimator. Also, we have optimized the decomposition process and added a selection phase that produces learning subsets of the same size [20] . Finally, we have also developed a parallel multi-threaded version of that decomposition/selection process.

Fault Tolerance

Application-level fault tolerance

Concerning fault tolerance, we have worked with Marc Sauget, from the University of Franche-Comté, on a parallel and robust algorithm for neural network learning in the context of the Neurad project  [35] . A short description of that project is given in Section  4.1.5 .

Our fault-tolerance strategy has shown to be rather efficient and robust in our different experiments performed with real data on a local cluster where faults were generated. Although those results are rather satisfying, we would like to investigate yet more reactive mechanisms as well as the insertion of robustness at the server level.

Programming model and frameworks for fault-tolerant applications

During the PhD thesis of Constantinos Makassikis [11] , supervised by Stéphane Vialle, we have designed a new fault tolerance programming model (MoLoToF) to ease the development of fault-tolerant distributed applications. Main features of MoLoToF include so-called “fault-tolerant skeletons” to embed checkpoint-based fault tolerance within applications, and enable various collaborations, such as application-semantic knowledge supplied by users to the underlying system (e.g.: middleware), in order to fine tune fault tolerance.

Two development frameworks have been designed according to two different parallel programming paradigms: ToMaWork for Master-Workers applications [17] and FT-GReLoSSS (FTG) for some kind of SPMD applications including inter-node communications [10] . The programmer's task is limited. He only needs to supply some computing routines (functions of the application), has to add some extra code to specify a fault-tolerant parallel programming skeletons and then to tune the checkpointing frequency.

Our experiments have exhibited limited runtime overheads when no failure occurs and acceptable runtime overheads in the worst case failures. Observed runtime overheads are less than the ones obtained with all other system-level fault tolerance solutions we have experimented, while maintaining very limited development time overhead. Moreover, detailed experiments up to 256 nodes of our cluster have shown that it is possible to finely tune the checkpointing policies of the frameworks in order to implement different fault tolerance strategies, for example, according to cluster reliability.

In 2011, we have used the FTG framework to make fault-tolerant an existing parallel financial application [31] from EDF R&D, where it is used for gas storage valuation. The resulting application kept its initial runtime performance despite some source code modifications which are required in order to use FTG. As it was the case in earlier experiments with other applications, these modifications accounted for a limited development time overhead and fault tolerance remained more efficient than system-level fault tolerance solutions.