Sorry, for the late entry - I completely missed setting a deadline reminder in my calendar
So many cores, so little scaling — Introduction to the massively-parallel solver PIPS-IPM++ and block-structure decomposition approaches
Format: Workshop (short presentation + hands-on)
Organiser: Manuel Wetzel
Preferred duration: 1.5h
Description: Energy system optimisation models (ESOMs) have become increasingly detailed in order to capture the complexity of modern energy systems. While the high spatial and temporal resolution enables more accurate analyses, it also leads to optimisation problems of enormous size. Even when linear formulations are used, solving these problems can require substantial computing time. Therefore, the standard approach today is to rely on highly optimised interior point methods provided by commercial vendors. While these methods perform well for many problems, they are limited in their ability to exploit the parallelism offered by today’s many-core computing systems. Better scalability usually requires users to manually implement decomposition techniques, which is time-consuming and requires expertise.
But what if decomposition and parallel solving could happen automatically?
This is precisely the question posed by the PEREGRINE project. We are developing methods to automatically detect exploitable model structures and utilise them for massively parallel optimisation with the open-source solver PIPS-IPM++. Our goal is to reduce the barriers to entry for advanced decomposition techniques, enabling model developers to get the benefits of high-performance parallel optimisation without having to redesign their models from scratch.
To this end, we created an expanding range of tools to help users at every stage of the process, no matter whether you want to better understand the structural properties of your optimisation model, experiment with user-defined decomposition strategies or evaluate fully automatic decomposition approaches. At the same time, we are making PIPS-IPM++ more accessible by developing new interfaces that simplify its integration into existing modelling workflows. Prototype interfaces for JuMP and Python are already available in a pre-release phase, making it easier than ever to test the solver with your own applications.
Are you curious to know whether your model could benefit from massively parallel optimisation? Bring your own optimisation model, explore its structure and find out if PIPS-IPM++ works for you. You may find that you can achieve significantly faster solution times with far less effort than you expected.
Requirements:
-
If you would like a better understanding of the underlying structure of your optimisation project, please prepare some problem files in either LP or MPS format.
-
If you would like to explore the automatic decomposition method, you will need access to a Linux environment (either Linux directly or a containerised solution such as WSL2 or Docker).
-
Access to a Linux environment is also required if you want to run the PIPS-IPM++ solver for your problems.
Optional links:
Repository for the tool for automatic annotation and visualization:
Repository for the PIPS-IPM++ solver: