How to use the AMD Threadripper for Scientific Computations

In the last blog entry we introduced a new optimization algorithm -  GCMA-ES   - and a new method to apply Lambert transfers for trajectory optimization - the Stochastic Lambert Solver. But we applied both methods only to variants of existing problems from the ESA GTOP optimization benchmark database GTOP Benchmarks .  As I showed Ingo Althöfer these results, he had the idea to apply the methods to something new. Ingo is Team-Manager of Team Jena which participated in all GTOC competitions since GTOC 5. This team has two special properties: it is both the smallest and the least successful team in the top ten. Probably these properties correlate. Beside its manager Team Jena has only one worker, the author of this blog. GTOC 10 is near, so if my manager has a training task for me, I should start working on it. And it is getting cold in German November 2018, but not in my flat as long as my AMD TR 2990WX is running full thrust utilizing all 64 threads. Around the W...

Open source real world optimization benchmarks hard enough to drive existing methods towards their limits are extremely rare. ESA’s Messenger full mission benchmark GTOP Task Messenger Full fulfills this criteria and therefore is currently a kind of reference problem to evaluate the effectiveness of optimization methods. It resembles an accurate model of Messengers full mission trajectory, so its solution is of real value for space mission planning. Currently the playing field is dominated by stochastic methods like ACO (Ant Colony Optimization), its commercial implementation Midaco (Mixed integer distributed ant colony optimization) and the open source CMA-ES (covariance matrix adaptation evolution strategy). In MIDACO Messenger simple retry mechanisms for ACO/Midaco and CMA-ES are compared – resulting in a slight advantage for ACO/Midaco over CMA-ES. Simple retry mechanisms generally are not able to solve the Messenger Full problem. Based on ACO/Midaco a more sophisticated retr...

Last time TR 2990 WX for scientific computing we showed how a 100 CPU Spark cluster may be replaced by a single CPU without any loss of performance for the Messenger-Full GTOP optimization problem. This time we will have a closer look at the GTOC1 problem GTOC1 and its approximation as part of the GTOP database. GTOC1 aims to remove a threat for planet earth caused by an asteroid headed in our direction. A small spaceship using an highly efficient low thrust ion engine is sent to hit the asteriod with so much force it will alter its course. “Force” means high impact speed at the asteroid at the “right” angle while maintaining as much mass as possible – which means use only a minimal amount of propellant. Planet gravity assist (GA) maneuvers help to achieve this goal. In the GTOP database there is simplified version of this problem GTOP - GTOC1 using a fixed planet flyby sequence approximating the low thrust engine fuel consumption by using the delta velocities (DVs) which need to to...

The AMD Threadripper 2990 WX is a processor suitable for scientific computations but since only two of its four NUMA nodes have direct memory access it is difficult to exploit its power completely. Its main advantages are price / thread - about 1800/64 = 28.125 $/thread power consumption / thread - 250/64 = 3.9 W/thread up to 350/64 = 5,4 W/thread overclocked These numbers look much worse if we don't utilize all 64 parallel threads, so we have to find a way to use them all thereby minimizing the penalty caused by the NUMA architecture of the 2990 WX. Overclocking / Power Efficiency Since power consumption grows super-linear with clock speed only limited overclocking (<= 3.8Ghz all cores) makes sense. In my experience using a fixed processor speed increases stability for specific VCore (I used VCore=1.25V for 3.8GHZ), but limits single core performance. Power consumption relative to performance is best at stock speed but can further be optimized lowering VCore (undervoltin...