Distribution of a Stochastic Control Algorithm Applied to Gas Storage Valuation

Abstract

International audienceThis paper introduces a research project that aims to speed-up and size-up some gas storage valuations, based on a Stochastic Dynamic Programming algorithm. Such valuations are typically needed by investment projects and yield prices of gas storage spaces and facilities. However, they involve computations which require great amounts of CPU power or memory. As a result, their parallelization on PC clusters or supercomputers becomes highly attractive and some-times unavoidable despite its complexity. Our parallelization strategy is based on a message passing paradigm, and distributes both computations and data on a cluster, in order to achieve speed-up and size-up. It includes some complex and optimized data exchanges which are dynamically computed, planned and achieved at each computation step. This optimized data distribution and memory management allows to process large problems on a high number of processors. More-over, our parallel implementation is able to support different price models, and our first experiments on a standard 32 PC cluster show very good performances particularly for complex price models

FT-GReLoSSS: A Skeletal-Based Approach towards Application Parallelization and Low-Overhead Fault Tolerance

“…Currently, the routing plan integrates two schemes based on the MPI_Issend-MPI_Irecv pair of MPI primitives. The first scheme performs all communications in parallel, while the second tries to "pace" them so that there is only a limited amount of ongoing concurrent communications at a time [18]. Concerning data movement, the routing plan groups sparse data into contiguous large messages and also minimizes copies from/to communication buffers by detecting ranges of contiguous elements.…”

“…The price models used at EDF for this software are a gaussian onefactor model (g), a normal inverse gaussian model (nig), and a two-factor gaussian model (g2d) [21]. The resolution method for the Swing application is the dynamic programming method that has been written in C++ and uses MPI for parallelization following the methodology in [18]. It also makes extensive use of the Blitz library for its convenient array manipulation facilities [22], and is about 18380 logical lines of code.…”

Large scale distribution of stochastic control algorithms for gas storage valuation

“…Our first implementation was based on a Python toplevel program calling MPI communication routines through the Python Pypar module [7]. This allowed users to easily tune the top-level program and run different distributed computations, but it was limited to the use of MPI_Bsend() routine and did not run on Blue Gene/L (which did not support Python), and we limited our experiments to a small 32 PC cluster.…”

A deep learning model for gas storage optimization

“…For standard references, see the Section 12.6 of Geman (2009), Section 5.3.4 of Fiorenzani et al (2012), and Holland (2007, 2008. Other references include, e.g., De Jong (2015), De Jong (2008, 2011), Safarov and Atkinson (2017), Cummins et al (2017), Carmona and Ludkovski (2010), Bjerksund et al (2011), Thompson et al (2009), Hénaff et al (2018), Malyscheff and Trafalis (2017), Jaillet et al (2004), Warin (2012) and Makassikis et al (2007). Much of the literature puts more emphasis on the modeling (and prediction) of gas prices rather than on developing algorithms for the optimization of storage plans.…”