Highly Scalable Full-Waveform Inversion on the Cloud Using Graphics Processing Units

Okita, Nicholas T.; Camargo, Alexandre William; Ribeiro, José; Benedicto, Caian; Coimbra, Tiago A.; Faccipieri, Jorge H.; Tygel, Martin

doi:10.3997/2214-4609.202011049

Cited by 1 publication

(2 citation statements)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…However, Okita et al . (2020) have shown that far too few shots in a single task can significantly decrease performance. Such a challenge also arises if we perform IC too often.…”

Section: High‐performance Computing Techniquesmentioning

confidence: 99%

See 1 more Smart Citation

High‐performance computing strategies for seismic‐imaging software on the cluster and cloud‐computing environments

Okita

Camargo

Ribeiro

et al. 2021

Geophysical Prospecting

View full text Add to dashboard Cite

Modern subsurface imaging techniques allow obtaining high‐quality images but with high computational costs. Nonetheless, depending on the amount of data, their execution is limited by memory in the current generation's hardware. However, with the advancement of new hardware and cloud‐based solutions, these problems are mitigated but still with the risk of work loss and instability. To mitigate the execution problems in memory‐limited and fail‐prone environments, we propose two high‐performance computing techniques. The first is based on independent checkpointing alongside a fault‐tolerant framework to store an execution state and recover from that state in case of failures. Besides, for memory‐limited graphics processing units, we present a technique to reduce the amount of memory requirement that we call the hybrid strategy. The experiments showed that the independent checkpointing alongside the fault‐tolerant framework is able to mitigate the performance penalty of node failures, with the independent checkpointing technique being more relevant when multiple nodes are terminated. Furthermore, the hybrid strategy technique has shown the possibility of execution of larger models that could make the graphics processing unit run out of memory otherwise. Finally, our implementation is scalable, allowing a significant improvement in performance when adding new nodes. In conclusion, our techniques can be used to deliver fast, high‐fidelity subsurface imaging in unstable and memory‐limited environments, such as the cloud.

show abstract

“…However, Okita et al . (2020) have shown that far too few shots in a single task can significantly decrease performance. Such a challenge also arises if we perform IC too often.…”

Section: High‐performance Computing Techniquesmentioning

confidence: 99%

“…Notably, it would be possible to reduce the task size to a single shot per node, thus disregarding the need for our IC process. However, Okita et al (2020) have shown that far too few shots in a single task can significantly decrease performance. Such a challenge also arises if we perform IC too often.…”

Section: Fault Tolerance and Checkpointingmentioning

confidence: 99%