Lossy checkpoint compression in full waveform inversion: a case study with ZFPv0.5.5 and the overthrust model

Abstract. This paper proposes a new method that combines checkpointing methods with error-controlled lossy compression for large-scale high-performance full-waveform inversion (FWI), an inverse problem commonly used in geophysical exploration. combination can significantly reduce data movement, allowing reduction run time as well peak memory. In the exascale computing era, frequent transfer (e.g., memory bandwidth, PCIe bandwidth GPUs, or network) is performance bottleneck rather than FLOPS of processing unit. Like many other adjoint-based optimization problems, FWI costly terms number floating-point operations, large footprint during backpropagation, and overheads. Past work adjoint has developed requirements backpropagation at cost additional computations. Combining this traditional compression, we explore three-way tradeoff between memory, precision, to solution. We investigate how approximation errors introduced by forward solution impact objective function gradient final inverted Empirical results from these numerical experiments indicate high lossy-compression rates (compression factors ranging up 100) have relatively minor on convergence quality