Current code does:

void ShenandoahHeap::ref_processing_init() {
  assert(_max_workers > 0, "Sanity");

  _ref_processor =
    new ReferenceProcessor(&_subject_to_discovery, // is_subject_to_discovery
                           ParallelRefProcEnabled, // MT processing
                           _max_workers, // Degree of MT processing
                           true, // MT discovery
                           _max_workers, // Degree of MT discovery
                           false, // Reference discovery is not atomic
                           NULL, // No closure, should be installed before use
                           true); // Scale worker threads

  shenandoah_assert_rp_isalive_not_installed();
}

...and:

  if (FLAG_IS_DEFAULT(ParallelRefProcEnabled)) {
    FLAG_SET_DEFAULT(ParallelRefProcEnabled, true);
  }

...so it always runs with parallel reference processing, even when -XX:ParallelGCThreads=1. This probably penalizes the single-threaded containers. Shenandoah should use single-threaded RP when that happens. Look for a relevant block in G1 code.