BPKSR: A Batch Parallel Krylov Subspace Recycling Method for Efficient Periodic Small-Signal Analysis in RF Circuit Simulation

Periodic small-signal analysis in RF circuit simulations poses significant computational challenges, particularly when dealing with a large quantity of frequency points. Although Krylov subspace recycling and parallel computing are effective strategies for accelerating this process, their synergistic combination has remained unexplored. We propose BPKSR, a novel parallel multi-thread subspace recycling method specifically designed for periodic AC (PAC) and periodic noise (PNoise) analysis. Initially, the frequency points are partitioned into multiple batches. For each batch, parallel processing of frequency points is carried out, leveraging the reuse of vectors in shared memory to minimize the number of iterations. Subsequently, a subspace is generated or augmented within each batch and then updated in the shared memory for recycling. Additionally, a frequency points allocation scheme is developed by jointly considering both the similarity between right-hand-side (RHS) vectors and the total number of threads, which aims to optimize parallel computational workloads.