Author: Robert Hyatt
Date: 15:04:29 08/25/03
Go up one level in this thread
On August 25, 2003 at 17:36:22, Sune Fischer wrote: >On August 25, 2003 at 17:28:33, Robert Hyatt wrote: > >>On August 25, 2003 at 16:28:31, Sune Fischer wrote: >> >>>On August 25, 2003 at 16:17:34, Dan Andersson wrote: >>> >>>> I would rather have 2.4 MNps on a single CPU :) And a single agent search will >>>>be more efficient except special cases. A pardox is that larger cache can be >>>>counter productive for some parallel applications due to increased memory >>>>traffic. >>> >>>Yes, but this memory trafic is a different animal than what the thread was >>>talking about. >>> >>>The tread was about making/uncopying, and for that you don't need inter chip >>>communication, each thread can run independently and can uncopy independently on >>>its own stack. For this double cache bandwidth is good, two chips can simply >>>copy double as fast as one chips so Bob's argument that 2.4 Mnps was more than >>>twice as bad 1 Mnps isn't valid concerning this specific issue. >> >>That is simply incorrect. >> >>Whether you have a dual or a single, if you run at 2.4M nodes per second >>you have a definite cache bandwidth requirement, and a definite memory >>bandwidth requirement, for my particular program, running at that particular >>NPS. >> >>IE if I can run at 2.4M nodes per second on a _single_ CPU, I _still_ have >>to copy _exactly_ the same number of bytes per node per copy/make. If I >>do them 1/2 as fast, but do them twice on two processors at the same time >>does not matter. The total bandwidth to support 2.4M nodes per second is >>a fixed number whether I do it with 1 processor or 1024. > >Bob please. >You compared 2.4 Mnps to 1 Mnps but you _can't_, such a comparison doesn't make >sense. > >If you are going to go 2.4 on _one_ cpu that cpu most likely will be 2.4 times >as fast and therefore have 2.4 times the cache bandwidth, so nothing changes >with respect to uncopying. >Yes you would need to copy 2.4 times as much, but you would also have an >equivalently larger bandwidth, status que! (status quo). That is exactly what I said. The issue is how many copy/makes have to be done per unit of time. Or how many units of time it takes to do a single copy/make. Depends on your time scale (1s or 1ns). A single cpu that will run crafty at 1M nps has a cache-cache and cache-memory bandwidth of X bytes/second. A single cpu that runs crafty at 2M nps has exactly twice the cache-cache and cache-memory bandwidth and twice the clock frequency. A dual-cpu just needs two cpus, but the two cpus give twice the cache-cache bandwidth, but _no_ improvement in cache-memory bandwidth. This was all about memory bandwidth with respect to copy/make. > >You made it sound like 2.4 was much worse than 1, which is just a silly >comparison at best. Clearly 2.4 _is_ much worse than 1. It requires 2.4X the bandwidth. It isn't something that can be ignored or waved away, obviously. Slow programs (IE the king) might not notice anything bad about copy/make. A program that is pretty fast (Crafty, for example) has to do it way more often. And it becomes a bigger piece of the total effort expended. That was why I used precisely 2.4M, to make it match to crafty on my specific hardware. It gives us _real_ numbers to work with, rather than hypothetical numbers that go with no real program at all. >Fact of the matter is 2.4 is worse than 1 because it takes place on a _dual_, >and that is the critical point I believe. No. 2.4 is worse because it has to do 2.4X as many copy/makes per second as a 1M nps program has to do. Bandwidth is definitely limited. The dual makes it even worse, of course. But version 9.15 used copy/make and version 9.16 used make/unmake. The speed gain was 25%... For Crafty, version 9.15 to 9.16. That's all I have said. It could easily be different for another program. If they are faster and copy more, it could be much worse. If they are much slower or copy much less stuff (or both) then it could be much less difference. But for crafty, the numbers are there... > >-S.
This page took 0.06 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.