Parallel Multithread Analysis of Extremely Large Simulation Traces

Abstract

With the explosion in the size of off-the-shelf integrated circuits and the advent of novel techniques related to failure modes, commercial Automatic Test Pattern Generator and fault simulation engines are often insufficient to measure the coverage of particular metrics. Consequently, a general working framework consists of storing simulation traces during the analysis phase and collecting test statistics from post-processing. Unfortunately, typical simulation traces can be hundreds of gigabytes long, and their analysis can require several days, even on large and powerful computational servers. In this paper, we propose a set of strategies to mitigate the evaluation time and the memory needed to analyze huge dump files stored in the standard Value Change Dump format. We concentrate on burn-in-related metrics that current commercial fault simulators and Automatic Test Pattern Generators cannot evaluate. We show how to divide the analysis process into several concurrent pipeline stages. We revise the logic process of each stage and all principal intermediate data structures, to adopt smart parallelization with very low contention and extremely low overhead. We exploit several low-level optimizations from modern programming techniques to reduce computation time and balance the different pipeline phases. We analyze simulation traces up to almost 250 GBytes computing different testing metrics. Overall, we can keep under control the memory usage, and we show time improvements of over two orders of magnitude compared to previously adopted state-of-the-art tools.

A Toolchain to Quantify Burn-In Stress Effectiveness on Large Automotive System-on-Chips

“…VCD-based methods (the lower path of Figure 5) rely on the standard VCD format to evaluate the BI metrics. They often run ad-hoc tools in a post-processing phase without the necessity of performing fault simulation [31].…”

“…The VCD File analyzer is a crucial tool within the toolchain, and it is based on the parallel analysis of typical data format [31].…”

“…The VCD analysis performs a transformation of the original VCD file from a time-based view to a signal-based view. Experimental results presented in [20], [31] are updated with VCD files ranging from a few GBytes to TBytes. The execution time bottleneck, as stated in [31], is always the file reading from the disk.…”

“…Experimental results presented in [20], [31] are updated with VCD files ranging from a few GBytes to TBytes. The execution time bottleneck, as stated in [31], is always the file reading from the disk. Table 4 shows single-point analysis elaboration times for each stress pattern; the size of input and output files is reported.…”