Re:Optimizing Indirect Memory References with milk
by forgotpwtomain on Wednesday September 14, 2016 @08:42AM
(#12496244)
> These press releases are terrible. It took me quite a while to find the original paper:
Was about to post the same-thing, how do these articles get past the editor? At any-rate it may be worth switching to the ACM publication link, the mit-news page is totally uninformative.
[Optimizing Cache Performance for Graph Analytics](https://arxiv.org/abs/1608.01362) also looks to be relevant (same ,authors Aug 2016).
Re:Optimizing Indirect Memory References with milk
by convolvatron on Wednesday September 14, 2016 @04:34PM
(#12500999)
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)
Re:Optimizing Indirect Memory References with milk
by _delirium on Wednesday September 14, 2016 @09:58AM
(#12496889)
It looks like the link here is trying to deep-link the PDF, which will work if you've already visited it "properly", but not if you haven't (haven't dug into what precisely it's checking, maybe a cookie, or your IP). You have to click through from here: http://dl.acm.org/citation.cfm?id=2967948
Re:Optimizing Indirect Memory References with milk
by unsignedqword on Wednesday September 14, 2016 @12:30PM
(#12498643)
Interesting, most people would say "the milk went sour" or "the milk soured," and not that it went "stale." "Stale" as a word just sounds weird in the context of a drink.
Optimizing Indirect Memory References with milk
http://dl.acm.org/citation.cfm?id=2967948
(seems to be available without paywall)
Was about to post the same-thing, how do these articles get past the editor? At any-rate it may be worth switching to the ACM publication link, the mit-news page is totally uninformative.
[Optimizing Cache Performance for Graph Analytics](https://arxiv.org/abs/1608.01362) also looks to be relevant (same ,authors Aug 2016).
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)