We want a thread-local GC for SICL. If most collections are thread-local, we would improve upon latency and throughput, even with many threads.
The plan for a while was to use something like the Doligez-Leroy-Gonthier (DLG) collector [fn:thread-local-ml]. However,