CONCURRENTLY EMPLOYING RESOURCES OF SEVERAL SUPERCOMPUTERS WITH PARASCIP SOLVER BY EVEREST PLATFORM

Sergey A Smirnov ,Vladimir V Voloshinov ,Oleg V Sukhoroslov

doi:10.54546/mlit.2021.34.36.002

Abstract

ParaSCIP is rather advanced open-source solver for discrete and global optimization problems. Thissolver is distinguished by that it can run on distributed memory systems and use up to 80,000 cores,solving open problems from the MIPLIB test libraries. Earlier, using this solver, we confirmed theconjecture on optimal packing of nine congruent circles on a square flat torus. The goal of the studywas to increase computing performance by utilizing resources of multiple clusters to solve hardoptimization problem. To do this, we use the previously developed DDBNB application, which allowsto speed up the solution of optimization problems by using coarse-grained parallelization based on astatic decomposition of feasible domain made before solving starts. DDBNB is an application for theEverest distributed computing platform which is responsible for running jobs on heterogeneouscomputing resources (servers, cloud instances, clusters, etc.). As a result, DDBNB, Everest, andParaSCIP had to be modified to make it possible to exchange incumbents (feasible solutions found bythe solver) between several ParaSCIP instances running on different supercomputers. The resultingsystem was benchmarked using three different instances of Traveling Salesman Problem. Thesupercomputers HPC5 of the NRC “Kurchatov Institute” and cHARISMa of the HSE University wereused as computing resources. As a result, for two problem instances, there is an effect, and thespeedup is especially noticeable for a more complex problem. However, for a simpler problem, theexchange of incumbents does not seem to affect the amount of speedup. For the third instance, there isno particular effect, at least no slowdown is observed.

Highlights

DDBNB [1, 2, 3] is a distributed application processing a set of computational tasks in parallel
DDBNB is an application for the Everest distributed computing platform [4], http://everest.distcomp.org, which is responsible for running jobs on heterogeneous computing resources
DDBNB, Everest, and ParaSCIP had to be modified to make it possible to exchange incumbent solutions between several ParaSCIP instances running on different supercomputers

Summary

Introduction

DDBNB [1, 2, 3] is a distributed application processing a set of computational tasks in parallel. The key features of DDBNB are: 1) immutability of the specified set of tasks; 2) the minimum data exchange between these tasks (only incumbent solution values) This is consistent with the coarse-grained parallelism model. ParaSCIP is a parallel application that runs on a computing cluster and uses MPI (Message Passing Interface) to exchange data between solver processes and a Load Coordinator process [fig. When solver finds a new incumbent solution, our code converts it to original problem coordinates, saves it to a file on disk and reports file’s name to a separate process which does data exchange with Everest server. When a new incumbent solution is received from outside, our code reads it, converts and sends it to ParaSCIP’s main process (Load Coordinator [fig. 1]) with MPI_Isend

Computational Experiments

Conclusion and Future Work