EnglishUništi kolačić - odaberi jezik prema postavkama web preglednika

Batched matrix operations on distributed GPUs with application in theoretical physics

Mijić, Nenad; Davidović, Davor (2022) Batched matrix operations on distributed GPUs with application in theoretical physics. In: 2022 45th Jubiilee International Convention on Information, Communication and Electronic Technology (MIPRO). pp. 293-299 .

[img]
Preview
PDF - Verzija prihvaćena za objavljivanje - članak
Download (318kB) | Preview

Sažetak

One of the most important and commonly used operations in many linear algebra functions is matrix-matrix multiplication (GEMM), which is also a key component in obtaining high performance of many scientific codes. It is a computationally intensive function requiring O(n3) operations, and its high computational intensity makes it well-suited to be significantly accelerated with GPUs. Today, many research problems require solving a very large number of relatively small GEMM operations that cannot utilise the entire GPU. To overcome this bottleneck, special functions have been developed that pack several GEMM operations into one and then compute them simultaneously on a GPU, which is called a batch operation. In this research work, we have proposed a different approach based on linking multiple GEMM operations to Message Passing Interface (MPI) processes and then binding multiple MPI processes to a single GPU. To increase GPU utilisation, more MPI processes (i.e. GEMM operations) are added. We implement and test this approach in the field of theoretical physics to compute entanglement properties through simulated annealing Monte Carlo simulation of quantum spin chains. For the specific use case, we were able to simulate a much larger spin system and achieve a speedup of up to 35× compared to the parallel CPU-only version.

Vrsta građe: Rad s konferencije/radionice objavljen u zborniku radova (NEODREĐENO)
Ključne riječi: matrix multiplication ; batched operations ; GPU ; MPI ; HPC
Područja: PRIRODNE ZNANOSTI > Matematika
PRIRODNE ZNANOSTI > Matematika > Algebra
TEHNIČKE ZNANOSTI > Računarstvo
TEHNIČKE ZNANOSTI > Računarstvo > Procesno računarstvo
Zavodi: Centar za informatiku i računarstvo
Projekti:
Naziv projektaVoditelj projektaŠifra projektaVrsta projekta
Skalabilni algoritmi visokih performansi za buduće heterogene distribuirane računalne sustaveDavidović, DavorUIP-2020-02-4559HRZZ
Napredne metode i tehnologije u znanosti o podatcima i kooperativnim sustavima - IJ za znanost o podatcimaLončarić, SvenKK.01.1.1.01.0009-1EK
Pohranio: Davor Davidović
Datum pohrane: 18 Aug 2022 09:23
URI: http://fulir.irb.hr/id/eprint/7514
DOI: 10.23919/mipro55190.2022.9803591

Actions (login required)

Pokaži zapis Pokaži zapis

Downloads

Downloads per month over past year

Contrast
Increase Font
Decrease Font
Dyslexic Font
Accessibility