Operating Systems, Cores and Work Units Hi All This situation is very interesting, and it raises far more questions than it answers. This is what I have found out so far. 1. Clients older than V6.34 can not get a4 work units. 2. The majority of people seem to be running V6.34 and so should be able to get a4 WUs. 3. People running V7 in windows seem to get all a4 WUs. 4. People running V6.34 in Linux seem to get all a3 WUs. 5. People running V6.34 in windows seem to get a mixture of a3 and a4 WUs. 6. Very few people are running V7 in Linux. Conjecture and speculation I do not know how the servers assign work units to different clients but it seems that Windows users are getting mostly a4 WUs and Linux users are getting mostly a3 WUs. It may not be that simple. Lots of people run older versions of Ubuntu for folding, I think 10.10 is popular, or use disc images from Linuxforge with all the tweaks already set up for best performance. V7 will not install in Ubuntu 10.10 as it is designed to run on 11.04 or newer. People run these older software configurations because they have some performance advantage, but they have painted themselves into a corner because these older configurations do not pick up the newer a4 work units. This situation is so widespread that there is a backlog of a4 work units and Stanford is introducing extra bonus on a4s from Monday 23rd July to give people a push. Conclusion It seems that peoples efforts to optimize folding for every last point per day is disrupting the project and has led to Stanford taking this rather odd action to try and balance things up. What to do If you are a Linux user then get an up to date version of your chosen distribution (it`s free after all). And install FAH V7 from the front page of Stanfords website. If you are a Windows user you probably don`t need to do anything. Does any of this make sense?
I run v6.34 clients across Win7 and various Ubuntu's, ranging from 10.04 to 12.04 - I've never seen an A4 core. Is the A4 core a non-smp core??
Hi DocJonz I have been running V7 on windows for over a month and I have never had an a3, only a4 work units. From Monday a4 work units are worth an extra 10%
I use V7 in Linux (Ubuntu 12.04 and Linux Mint 12) they both get the A4 WUs all the time. Has anyone tried V7 in Linux with a 2P or 4P set up (thinking about bigadv or bigbeta)? I will probably test my 4P rig (when it's built ) with V7 standard SMP units as they seem to complete so quickly. Don't know whether the latest version of the Kraken works with A4 units? Any idea if it is even necessary when using a more modern distro than 10.10 ? More questions than answers I'm afraid
That is a very good question. I am finding a very big variation in PPD between work units, The best get about 45,000 PPD and the worst get about 7,000 PPD on my system.
quite a variation then my Q6600 at 3.6 gets around 8-9k ppd with the gtx 560 running i wonder what it will get on the A4 core i will move to v7 and find out
I'm running 3 machines each with Win 7-64bit, V7 clients and Sandybridge 2600k processors. I've seen PPD vary between 7.5k and 53k There was definitely a boost to my daily totals when I upgraded from v6 clients to v7 so that's probably still worth doing whatever OS you're using.
The Kraken, NUMA & SMP Hi One Box Short answer Use the Kraken on multi-socket systems. Long answer The Folding at Home SMP cores are designed to run on an SMP system. SMP stands for Symmetric MultiProcessing, a group of processors connected to a single memory bus so that all processors can access the same memory. It is normal for threads to be rotated through the physical cores to even out the load. This is not a problem because the thread can access its data in memory from any of the cores. You can have multiple memory channels to increase bandwidth but it is still a single block of memory. SMP works very well but if you want to build a computer with lots of processing cores then the memory bus will become a bottleneck in the system. Servers use NUMA, this stands for Non Uniform Memory Access. In this architecture the computer has multiple processing nodes and each node can have multiple cores. Each node has its own memory controller its own local memory and is connected to other nodes in the system by high bandwidth point to point links. NUMA enables you to build machines with a very large number of cores and enough memory bandwidth to support them. This is already quite complicated but it gets worse. Each node has a limited number of links so in a large system one node can not be connected directly to all the other nodes. This means that the data a thread requires could be in local memory, or in memory on an adjacent node (1 jump away), or in memory on a distant node (2 or more jumps away). With me so far? Remember that SMP threads jump from core to core to even out the load on each physical core. Well on a NUMA system the data gets left behind on one node while the thread is now running on a different node, when it needs the data the thread has to send out a request and then wait for the data to catch up. Software designed for NUMA systems is coded to minimize these problems but FAH is designed for SMP systems. One solution is to use node interleaving, this splits up the data from each thread and stores it across all the nodes. It sounds like madness, but by evening out the delays the overall performance is faster. The Kraken binds each thread to a physical core so that waiting for data is minimized. This is by far the best solution. When using The Kraken turn node interleaving off. To give you an idea of how complicated this gets, Magny-Cours chips each contain 2 nodes. So a 4P 6128 system has 8 nodes with 4 cores each and each node has 3 links. To illustrate the interconnections draw 8 nodes in a circle, link each node to its neighbour on both sides to form an octagon and then link each node to the one directly opposite. I said it was a long answer.
just started the V7 client on the Q6600 lets see how it goes.... and i will install the V7 client on my P4 system see how that goes, just need to remember to add 10% to the PPD on the A4 core WU's as it wont show in HFM etc does anyone know have to uninstall the V6.34 client from Ubuntu? or is it just a case of removing the files? thanks
@ Slowlemon Thanks for the very detailed explanation, I will remember to use the Kraken when my 4P sytem is up and running. (Hope to have the mobo Friday).
Hhmmmhh... not pleased currently with V7. Just switched over a few days ago, using V6.34 I got a pretty consistent PPD but now I am getting 34k PPD or 44k PPD for some units then it drops all the way to 4k PPD for a couple. These A4 units are seriously screwed and I am not sure if my average PPD will increase because of this or decrease.
So far ive had a couple of A4 WU's P8042 and I'm getting 10200 ppd plus 10% bonus on top of that. The A3 WU's where 8-9k ppd That's on my Q6600, I will change the 4P rig tomorrow when it finishes the big beta WU
Mine's been folding for 24 hours and recorded no points, so I switched back over to 6.34 before work. I'll have a look why later...
43,000 PPD on a 2500k running at 4.4Ghz, and that's without including the 10% bonus. Some of these A4 units are compatible to the scores you used to get when spoofing bigadv.