Details

Autor(en) / Beteiligte

• Shan, Ruiwen
• Calhoun, Jon C.

Titel

Exploring Data Corruption Inside SZ

Ist Teil von

• 2022 IEEE International Conference on Big Data (Big Data), 2022, p.3172-3178

Ort / Verlag

IEEE

Erscheinungsjahr

2022

Quelle

IEEE Electronic Library (IEL)

Beschreibungen/Notizen

• Due to the increasing scale of scientific research, scientists need to collect massive amounts of data to solve complex scientific problems. The exponential growth of data poses significant challenges to high-performance computing (HPC) systems in terms of their computational ability, storage capacity, and transmission bandwidth. Data reduction techniques such as data compression have become one of the most promising solutions to these problems. Error-bounded lossy compression is now commonly utilized in HPC systems to substantially reduce data volume while precisely maintaining data accuracy. However, the majority of research was done on improving compression efficiency, such as compression ratio, and insufficient attention is paid to the security of the compression process.In this paper, we concentrate on the impact of corruption on error-bounded lossy compressor SZ, including corruption due to transient failures of hardware and corruption injected by malicious users. We analyze and quantify the influence of this corruption on compressed datasets by simulating the corruption errors that occur in the regression coefficient values and computation during compression using four failure models. The results demonstrate that SZ's prediction-based design makes it sensitive to corruption of the regression coefficients. A single bit-flip in the regression coefficients can result in noticeable error propagation, in some cases, the compression ratio fluctuates up to 0.28%, but peak signal-to-noise ratio(PSNR) drops to negative levels.