Author: Michael Nieves
Date: 22:55:05 10/01/01
Go up one level in this thread
On September 30, 2001 at 14:34:23, Slater Wold wrote: >On September 30, 2001 at 11:42:28, Joshua Lee wrote: > >>On September 29, 2001 at 15:34:45, Slater Wold wrote: >> >>>On September 29, 2001 at 14:42:02, Joshua Lee wrote: >>> >>>>Sue me if i am wrong but you don't get 2X with Dual Systems more like 1.7x and >>>>sometimes less sometimes more when in comes to chess so 2380Mhz sounds more >>>>correct. >>> >>>I won't sue you. :) >>> >>>But you are correct. Some get 1.9x (Crafty) and some get 1.3x (Deep Fritz). >>> >>>I think it's a little more than 2380mhz though. Perhaps 2600mhz. Which is >>>still faster than 1600mhz. >>> >>>Not to mention, I can get a full 1400mhz out of 2 applications, while he can >>>only get 1600mhz out of 1 application. >>> >>> >>>Slate >> >>How do applications like SETI , Distributed net, etc run with the duals? >> >>I have tried SETI with my now anchient 800Mhz athlon and it takes upto 2 days >>for a seti block to finish. If Cancerbusters works with Dual processor systems >>that would be great too. > >I know for a fact that their are distributed programs that are SMP. Not sure if >any of these are. > >I used to have the cancer screensaver back when I had my 2x1000mhz. And it used >200% of the CPU's. > > > >Slate I'd be very surprised if you did. You will never (he says) get full use of additional CPUs (in an SMP system) for an application. MPP is a different story (e.g. Deep Blue was on an MPP IBM SP2 with a high-speed switch). The 80% increase for a second CPU that the developers of Deep Fritz give are theoretical curves. In practice it is generally less. You'll get even less from 3rd and 4th CPUs (e.g. 3:+55%, 4:+30%, etc). This degredation curve will continue even to the point where you will decrease the performance of the whole system by adding additional CPUs! Your local perfmon won't notice a thing (it would be like solving a zen puzzle if it could). The real rub is that the faster the CPUs, the faster the inflection point is reached. This is a inherent problem with SMP (symmetrical Multi-proc) systems and the reason why computationally heavy apps (data-warehousing, CFDs, chess) are quick to go MPP when money is on the line. But MPP-able apps are much more complex and expensive. Why the SMP limits? In a nutshell: Each additional CPU is sharing resources with its companion CPUs (memory, disks, etc). The overhead involved in this management increases by almost an order of magnitude with each additional CPU. The curve is very dependant on the speed of communications between the CPUs (e.g. cross-bars) which is why faster CPUs degrade faster; they gag the pipeline quicker. It also explains why some apparently identical SMP systems have big price differences: Fast cross-bars are expensive (you get what you pay for). MPP solves this by not sharing any resource (in-theory) and shifting complexity to the app. The SP2 (Deep Blue) uses an external, point-point high speed switch. Very well done when first introduced. In theory, you get 100% of each additional CPU. That is why I'd generally stick to Highest-end single CPU systems rather than dual systems with slightly slower CPUs. Be happy to discuss this further off-line.... Michael Nieves
This page took 0 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.