the press release is terrible. the paper is fine. this is really interesting work. of course the paper gives a much more nuanced view of the performance benefit

the press release is terrible. the paper is fine. this is really interesting work. of course the paper gives a much more nuanced view of the performance benefits. this is effectively scheduling to hide memory latency (ala SMT) except done entirely dynamically, with batching to support locality. the most impressive part of the paper isn't just the speedup, but the speedup in the presence of the substantial overhead involved in capturing, scheduling, and restoring the closures. it would be nice if the super-fine-grained threading model made more than the occasional appearance. it opens up a lot of potential for hiding all kinds of latencies. if a runtime can reap substantial benefits, imagine what an architecture could do (i.e. SMT)