Molecular Dynamic Simulation on High Performance Infrastrucutres

Diplomsko delo zajema primerjavo različnih računalniških platform visoko zmogljivega računalništva pri simulacijah molekulske dinamike, ki spada med računsko zelo zahtevne probleme in potrebuje veliko procesorsko moč. Naš namen je bil kritično ovrednotiti različne platforme pri reševanju molekulske dinamike, zato smo simulacije izvajali od eno jedrnega do 16 jedrnega procesorja na računalniški gruči in z uporabo ene in dveh grafično procesnih enot (GPE). Rezultati bodo služili kot pomagalo pri načrtovanju nabave nove strojne opreme za izvajanje biomolekularnih simulacij. Kot merilo smo uporabili čas, ki je potreben za izvedbo simulacije, ter kako dobro so platforme skalabilne. Pri primerjavi hitrosti se v biomolekularnih simulacijah uporablja mera ns/dan, ki nam pove, koliko nanosekund simulacije je sistem sposoben izračunati v enem dnevu. Za potrebe diplomskega dela smo simulirali velik hidratiran protein MAO B z endogenim substratom feniletilaminom. Simulirani sistem je izjemno relevanten za področje nevroznanosti, saj ta encim regulira nivoje biogenih aminov, ki imajo ključno vlogo pri prenosu živčnega signala. Ugotovili smo, da se simulacija molekulske dinamike občutno hitreje izvaja na GPE kot na klasičnih procesorjih. Poleg tega je tudi cenovno najbolj ugodna platforma za izvajanje klasičnih molekulskih simulacij. Z vidika skalabilnosti pa je smiselna uporaba le ene GPE hkrati, saj je pohitritev ob uporabi dveh GPE v primerjavi z eno slaba.This thesis covers comparison between different computer platforms of high performance computing while performing molecular dynamics simulations, which falls under very complex problems and needs lots of processing power. Our goal was to critically evaluate different platforms while solving molecular dynamics, so we used 1 to 16 processor cores on a computer cluster and one and two graphics processing units (GPU) for simulations. The results will be used while planning on buying new computer hardware for biomolecular simulations. We used time needed for simulations and platform scalability as our benchmarks. For comparing speed in biomolecular simulations we used ns/day for comparison. Ns/day tells us how many nanoseconds is a system capable of simulating in one day. For this thesis we simulated a large hydrated MAO B protein with endogen phenylethylamine substrate. The simulated system is extremely important for neuroscience, since it regulates levels of biogenic amines, which have an essential part in neuro signal transmitting. The results have shown us that the use of GPUs is significantly faster than regular processors when it comes to molecular dynamics. Moreover it is also the most cost effective platform for classical molecular dynamics. From the perspective of scalability it makes sense to only use one GPU at the time, since the speed-up when using two GPUs is lower than expected