This optimization is based on this paper [1]. The idea is to keep last successfully stolen queue as one of best of two candidates.

Thomas Schatzl and I did a few benchmark/performance tests on prototypes, all yielded positive results. Please see the email thread on shenandoah mailing list [2] for details.

[1] Characterizing and Optimizing Hotspot Parallel Garbage
    Collection on Multicore Systems
    http://ranger.uta.edu/~jrao/papers/EuroSys18.pdf

[2] http://mail.openjdk.java.net/pipermail/shenandoah-dev/2018-June/006502.html