Author: Sune Fischer
Date: 11:25:17 08/25/03
Go up one level in this thread
On August 25, 2003 at 13:44:13, Robert Hyatt wrote: >Cache copying is _important_. If I write to a line of cache on processor >A, then processor B has to invalidate that line of cache if it has it. That >hurts. Also when I write to a line of cache, that displaces what was there >_before_ the copy. And eventually that modified line of cache has to also >be written back to real memory. > >IE for copy/make, you generate a _lot_ of memory traffic... And with >two processors sharing a single memory bus, the traffic for two is the same >as for the traffic for one CPU that is twice as fast. > >Remember, cache-to-cache is a misnomer. You have to replace something that >is already there, and once you write to a cache line, that cache line _will_ >be written to memory before it is destroyed. So cache-to-cache in the case >of copy/make often turns into memory-to-memory. I'm not disagreeing with that, but you have also changed the topic a little bit. Your calculation was based on 2.4 Mnps and you compared that to 1 Mnps (nps being the _only_ factor changing in the equation) but you neglected to include the double bandwith _for the cache copying_ which would unify it back down. Somehow you changed to saying that using a double stack would trash the cache and require you to go over the bus, at least that is how I'm reading it. This may be true, or it may not be, but only experiment can tell us that and it will depend on the program. I could just as well assume that it wouldn't happen, and as a result your numbers were a factor 2 off, that is _all_ I was pointing out. Hoped this cleared it up :) >>Each thread would copy to it's own stack, which I presume would reside in the >>cache on each respective chip. >>Otherwise you're right it would be dead slow, but then that would be the same >>for all kinds of tables which I certainly hope isn't the case. > >It's a problem _everywhere_. And with a dual, it gets worse as a copy to a >line in one cache invalidates that line in another cache when the content >address is the same. That makes duals potentially much worse. Potentially, yes. Anyhow unmaking is faster for me even on single cpus, and I suspect you are right that uncopying would only get worse with more chips. -S.
This page took 0.08 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.