View Single Post
  #2 (permalink)  
Old 23rd June, 2010, 03:29 AM
ThunderRd's Avatar
ThunderRd ThunderRd is offline
Irreverent Query Chairman
Join Date: June 2007
Location: NYC native in northern Thailand
Posts: 2,241

The older clients didn't scale well when used on dual and quad core processors, so they didn't completely use the CPU cycles. There was a movement to install and run two clients. This used the CPU completely, and more ppd resulted. I tried it, too. It did produce more points, but...

It had a drawback, and that was that it actually hurt the science. Because two clients were running, each client [individually]ran more slowly than expected. Even though the box was producing more ppd, the results were being returned to Stanford slower, delaying the generation of the next WUs to be processed. For this reason, Stanford appealed to all users to run one standard client on each physical core, or one SMP client on each dual or quad setup. This insured that the results would be returned as quickly as possible.

People didn't listen, of course. They ran multiple clients for the points, and many ignored Stanford's wishes. We had a long thread on it here, as well. Here it is, it's posted in the edit at the top of the sticky thread you read. The thread is about the Linux client, but the principles are the same:
Dual Linux SMP clients on Quadcore

It was a heated debate for quite a while in the folding community forum, and lots of people were upset that they had invested in new processors, and they weren't being fully used.

This was why Affinity Changer was born, because it gave the user some flexibility as to how his CPU cores were being assigned.

Anyway, the bottom line is that the newer cores don't have the same low CPU usage, so there is very little, if any, reason to run two simultaneous clients on a quad anymore. The new cores use pretty near all of the cycles, and the WUs are returned as quickly as possible that way. There may be some very small gain in ppd, but not enough to make it worthwhile these days.

The quick return is an essential part of F@H that some people don't consider. Imagine that each project is like a pyramid. The lowest layer, the base, of the pyramid is generated first - so you get WUs from project x, generation 0 first. In order to produce the WUs in project x, generation 1, they need the results from generation 0. If there are some data missing from gen 0, not all of the WUs for gen 1 can be created. Same for gen 1 -> gen 2. The scientists need the completed data so they can build the next layer of the pyramid. Hence the new bonus point system for quick return. You get rewarded for more powerful hardware, running the box for more hours, stable overclocking, etc. All of these factors enable them to get the science more quickly, and your reward is the points.

The 5.9x clients are scheduled to be deprecated next month. I decided to leave the sticky post on dual client setup until 5.9 is gone. I'll move it to the deprecated information at that time.

Sorry for being long-winded, but this topic comes up fairly regularly, so it may be a good idea to have a full explanation of why it's not recommended, rather than me saying "don't do that"
#1: Tt Armor, ASUS Maximus Extreme, QX9650@4.1G, 8G Corsair Dominator GT DDR3-2000, Corsair HX1050, H2O-Swiftech, Gigabyte GTX470/Arctic Accelero Xtreme Plus II, Intel 520 SSD, Kingston SSD, 2xRaptor 150G RAID0, Win 7 Pro 64
#2: Tt Shark, ASUS P5Q Pro Turbo, Q6600@3.8G, 4G HyperX-1600, Corsair HX850, CoolerMaster V10, 2xASUS 9600GT, 2xRaptor 74G RAID0, OCZ Vertex 4 SSD, Gentoo/siduction Linux [64-bit]
#3, #4: Opteron 170@2.75G nude, A8N-SLI Deluxe, Gentoo

AOA Folding @HomeOur sister site:

Last edited by ThunderRd; 23rd June, 2010 at 03:33 AM.
Reply With Quote