3 ms·
Looking at the underlying paper, I'm actually skeptical that there's any practical performance/energy boosts from this accelerator. The best case they're giving
by jcranmer 8y ago
Looking at the underlying paper, I'm actually skeptical that there's any practical performance/energy boosts from this accelerator. The best case they're giving is something like a ~3-5× boost on GC (which is generally ~10% of total time, much of it hopefully idle time anyways if you're tuning GC well), with a 15% power reduction, assuming your collector is a stop-the-world, non-concurrent collector (and you rewrite your object layout). There's no measurement of the energy cost of the idling core, nor of the cost of extra memory and cache traffic trundling objects to and from the accelerator rather than sticking on the core. Those costs are likely to be even worse if you try to adapt this into incremental and concurrent collectors.
Especially when you ask the question "is it better to spend the area/power budget on a highly-specialized accelerator, or to give it over to another full core that can do other things when it's not collecting garbage," it's hard to motivate this kind of accelerator.