Folding Forum

Posted: **Thu Apr 26, 2012 1:24 am**

Hi guys,

I have been granted authorization by my college's vice-director and the IT/Computer Lab director (who very much sympathizes with the project and is helping me alot) to run FAH on their 50 computers with "AMD Athlon 64 X2 5600+" processor and 1 GB RAM each. These are dual-core 2.8 GHz processors. The computers are basically on 24/7, and most of the time they are not being used, so they stay in the log-on screen. I have begun testing on two of those by installing FAH as a service, and it succesfully calculates even when no one is logged on.
I have set up both as SMP:2, both WUs have a deadline of 25 days and the ETA for completion is ~3 days, so SMP seems a good choice, as there is more than enough time to take care of problems should anything go wrong. Uniprocessor slots don't seem to have this luxury as when I tested them the deadline was 4 days. However since 50 computers is a lot I want to make sure I'm using the best set-up I possibly can. So:

1- Do you guys think double uniprocessor slots rather than a SMP:2 slot are a better idea for computers of these configs?
2- Should I set the client type to advanced, thus getting advanced WUs? I heard they use more RAM, and 1 GB of RAM sounds a bit puny (the computers' disk usage orange lights already flash like crazy, indicating the 1 GB is probably not enough currently and it is using the HD).
3- My mom, who also has low RAM on her PC, was running FAH but she stopped it, claiming it made her daily backups take forever. I figured it was due to the disk usage. Am I correct? Do I have to ask the director if they have any routines which use the HD intensely? I'd prefer not to do this as he might back out if the performance of such routines would decrease :/

Posted: **Thu Apr 26, 2012 5:02 am**

1. SMP is always better than running multiple uniprocessor slots, just make sure the computer can finish the SMP WU before the preferred deadline
2. The flashing disk lights is most likely the hard drive space, not RAM, low RAM will usually just make the computer very slow with no indicators (not 100% sure on this though)
3. You can always check the memory usage by opening task manager

Posted: **Thu Apr 26, 2012 7:37 am**

2) Running a backup or a virus scan while FAH is working hard can cause performance issues. It's much less of a problem if a person is using the computer for "normal" things, particularly on a machine with low RAM. Running too many active tasks at the same time tends to cause thrashing. This is a bigger problem for a backup than for most other applications because the backup runs continuously and makes heavy use of the disk so it uses a (very) little bit of processing time and then relinquishes the processor to FAH which claims a bigger amount of time. In fact, many backup programs are specifically designed so they slow down when something else is running so it doesn't create noticeable lags from running continuously.

I expect that for an X2, running multiple uniprocessor slots/clients will produce fewer FAH points than smp:2 but will also have a lesser impact the other tasks. You'll need to evaluate it carefully.

Posted: **Tue May 01, 2012 8:29 pm**

bruce wrote:I expect that for an X2, running multiple uniprocessor slots/clients will produce fewer FAH points than smp:2 but will also have a lesser impact the other tasks. You'll need to evaluate it carefully.

Hmm, in terms of processor usage they should be the same, both cores will be running at 100% regardless, so the impact should be the same unless SMP uses more RAM or whatnot.

Posted: **Tue May 01, 2012 11:00 pm**

Yes, but 100% of a single core has to work harder to process two low priority tasks than to process one low priority task and leave the other one temporarily dormant.

Processing a pair of uniprocessor assignments are two logically independent processes. Say some other task puts demands on one CPU. One of the WUs will get paged out. It can be paged back in whenever there are resources available to work on it . . . at the convenience of the OS.. There is no need to sync the two processes.

A single SMP project is directly dependent on processing equal amounts of work in each of the two threads and requires constant synchronization. Say one thread gets paged out. The other one will (almost) immediately wait for the other one to catch up so the active one will be paged out and the dormant one will be paged in. The two will try to share the same resources by constantly switching to/from virtual memory.

It's not visible at a human time-frame, but the OS does care.

Posted: **Wed May 02, 2012 1:34 am**

Nantes wrote: 1- Do you guys think double uniprocessor slots rather than a SMP:2 slot are a better idea for computers of these configs?

Dunno. Setup two machines, one with uni, one smp; let them churn for a while; then check which configuration finishes WUs sooner and what PPD they give you. Earlier finish = more scientific value.

Nantes wrote: 2- Should I set the client type to advanced, thus getting advanced WUs? I heard they use more RAM, and 1 GB of RAM sounds a bit puny (the computers' disk usage orange lights already flash like crazy, indicating the 1 GB is probably not enough currently and it is using the HD).

I have two SMP:4 machines running, both setup as client-type:advanced, and in both cases they use barely a few megabytes of ram (< 20); 1GB is more than enough to run FAH. From v7 slot options pictorial guide:

Code: Select all

client-type       advanced       Assigns WUs in late-stage beta if available

It would seem you merely get certain WUs sooner than "normal" clients. Whether those WUs will use more ram, depends on the project; but i think FAH folks try to make the "normal" client usable for 99% of PCs.

OTOH, if you plan to use GPU clients, that's a different story. On my machine with ATI card, GPU client uses ~380M of ram (not video ram).

Anyway, if disk usage is going crazy, perhaps you have a real problem you should look into first. You should know that running FAH 24/7 is a serious stress test; your CPU will constantly generate heat, thus your pc case needs to have good cooling (and in your case also the room where all the PCs are

) Depending on your pc case, CPU heat can heat up the harddrive a bit, and then putting load on the harddrive could easily cross the limit.

One more tip - i keep my SMP:4 machine downclocked from 2.66 to 2.21. This gives me 17% less performance for about 40% less power and 40% less heat. Plus it's almost as quiet as when idling. If you find a way to underclock+undervolt even by a small bit, it's definitely worth it

Nantes wrote: 3- My mom, who also has low RAM on her PC, was running FAH but she stopped it, claiming it made her daily backups take forever. I figured it was due to the disk usage. Am I correct? Do I have to ask the director if they have any routines which use the HD intensely? I'd prefer not to do this as he might back out if the performance of such routines would decrease :/

I highly doubt FAH has caused her backup to slowdown.

First of all, the cores launched by FAH run at lowest possible process priority on both linux and windows. That means all applications with normal priority will get CPU time before FAH; FAH just uses the time your CPU would spend idling. For example, if you launch video encoding on all cores, FAH will get nearly no CPU time at all.

Second, FAH uses the harddrive very rarely - only when it writes checkpoints (default is every 30 minutes, IIRC)

Cheers

Posted: **Wed May 02, 2012 2:22 am**

nyanthiss wrote:
Nantes wrote: 1- Do you guys think double uniprocessor slots rather than a SMP:2 slot are a better idea for computers of these configs?
Dunno. Setup two machines, one with uni, one smp; let them churn for a while; then check which configuration finishes WUs sooner and what PPD they give you. Earlier finish = more scientific value.

That's true but it's a bit of an oversimplification. Both Uniprocessor and SMP WUs come from various projects that represent different time segments of different proteins. Thus they're assigned different deadlines and different numbers of points per WU. A WU which is completed in a quarter of the Deadline is generally worth more scientifically speaking, than a WU that is completed in half of the Deadline. A direct comparison between two different projects is meaningless unless the WUs happen to be from the same project.

Second, FAH uses the harddrive very rarely - only when it writes checkpoints (default is every 30 minutes, IIRC)

Again, mostly true. The default checkpoint interval is 15 minutes but you can adjust it to as long as 30 minutes if you choose.

Posted: **Fri May 04, 2012 12:23 pm**

Hmm, i stumbled into SMP FAQ -> Notes for running, saying

1. We strongly suggest people run this client on 4-core boxes. While it will run on 2-core boxes, we have noticed some potential problems (we are looking into these issues now).

Does anyone know how old that information from FAQ is, and what the "potential" problems are

? Page footer says "Last Updated on December 11, 2010"...

Posted: **Fri May 04, 2012 1:53 pm**

Old. The original "potential" problems are fixed.

But as shown above, folding SMP with only 2 cores can be problematic if there are other work loads on the CPU.

The recommendation to fold SMP with 4 (or more) cores still applies... So should the FAQ be updated or not?

Posted: **Fri May 04, 2012 4:27 pm**

7im wrote:The recommendation to fold SMP with 4 (or more) cores still applies... So should the FAQ be updated or not?

Yes it should, why keep the "potential problems" warning there if it's fixed already?

It might scare away potential dedicated 2-core folders

Posted: **Fri May 04, 2012 4:36 pm**

Because there are still potential problems, just newer ones... so you see absolutely no problems folding on a 2 core box?

Posted: **Fri May 04, 2012 4:39 pm**

If it can complete WUs before the timeout, why not?

Posted: **Fri May 04, 2012 6:20 pm**

Big IF, which continues to be a problem. Not all dual cores can meet the deadlines.

Just because you haven't seen any problems doesn't mean problems don't exist...

Rather than debate whether it should be updated or not, why don't you suggest new wording for that entry in the FAQ?

Posted: **Fri May 04, 2012 6:37 pm**

7im wrote:Big IF, which continues to be a problem. Not all dual cores can meet the deadlines.

I'm wondering if it would be possible to test the minimum hardware requirements (kinda like a game's system requirements)

7im wrote:Just because you haven't seen any problems doesn't mean problems don't exist...

But then this statement applies to all types of slots running on all types of hardware, right?

7im wrote:Rather than debate whether it should be updated or not, why don't you suggest new wording for that entry in the FAQ?

We recommend people run this client on 4-core boxes. While it will run on 2-core boxes, it may not be able to finish work units before the deadline. We suggest that people running this client make sure that it can finish work units before the preferred deadline, as it's scientifically much more valuable that way. If unable to finish work units before the preferred deadline and/or just making the final deadline, we suggest switching to one or more instances of the uniprocessor client.

How about that?

Posted: **Fri May 04, 2012 6:49 pm**

iceman1992 wrote:We recommend people run this client on 4-core boxes. While it will run on 2-core boxes, it may not be able to finish work units before the deadline. We suggest that people running this client make sure that it can finish work units before the preferred deadline, as it's scientifically much more valuable that way. If unable to finish work units before the preferred deadline and/or just making the final deadline, we suggest switching to one or more instances of the uniprocessor client.

How about that?

We recommend people run SMP projects on CPUs with 3 or more cores. While it will run on 2-core boxes or single core HyperThreaded CPUs, . . .

Folding Forum

50 dual-core 2.8 GHz computers: Most effective configs?

50 dual-core 2.8 GHz computers: Most effective configs?

Re: 50 dual-core 2.8 GHz computers: Most effective configs?

Re: 50 dual-core 2.8 GHz computers: Most effective configs?

Re: 50 dual-core 2.8 GHz computers: Most effective configs?

Re: 50 dual-core 2.8 GHz computers: Most effective configs?

Re: 50 dual-core 2.8 GHz computers: Most effective configs?

Re: 50 dual-core 2.8 GHz computers: Most effective configs?

Re: 50 dual-core 2.8 GHz computers: Most effective configs?

Re: 50 dual-core 2.8 GHz computers: Most effective configs?

Re: 50 dual-core 2.8 GHz computers: Most effective configs?

Re: 50 dual-core 2.8 GHz computers: Most effective configs?

Re: 50 dual-core 2.8 GHz computers: Most effective configs?

Re: 50 dual-core 2.8 GHz computers: Most effective configs?

Re: 50 dual-core 2.8 GHz computers: Most effective configs?

Re: 50 dual-core 2.8 GHz computers: Most effective configs?