Sun x4450....

[{PMS}fishy]

http://fishy.neu.edu/x4450.png

Have 2. Need to configure and burn in. Running RHEL5, what is the best config for PPD.

:D

[Adak]

I have a 4 processor Xeon 5420 * two (8 core) box that is a dedicated folder, atm.

I thought that having two SMP clients, with their 8 threads, would be the perfect balance for the 8 cores, but Charles said "use 3 SMP" for more ppd.

Charles was right, and I now fold 3 instances of the 5.91 beta 6 SMP client, with no affinity set, and no affinity balancer program.

OS is Windows XP, which recognizes about 3 Gig's of RAM, only. Each instance is told to report no more than 760 Megs of RAM, in advanced options. (512Megs. wasn't always quite enough).

What you have to watch out for is, that if you stop one instance of the SMP client, the other two *frequently* will lose all their work, and stop, as well. :(

So I never stop any instance of SMP, like normal (with Ctrl + c). Instead, I just cut off the PSU power switch, and stop everything, instantly. Sounds odd, but the few times I've done this, it has always worked. No work was lost.

In your place, I'd start with 4 instances of 5.91 SMP, and get a "baseline" ppd read from FahMon (having it set to only use the last 3 frames for it's calculations, through it's preferences). Then keep adding another instance, until the system began to decrease in ppd, always giving FahMon a good 90 minutes or so, to get a good read on the new setup.

I believe 6 instances will be close to being right, if your memory and cache will hold up OK.

FahMon: http://fahmon.net/
Newer versions are much better. :)

Let us know how it goes, please! That's an amazing setup! :beer: X 16

[{PMS}fishy]

2 things.

1. I wasn't very clear with what this is or what I am doing. The machine runs VMWare ESX, so I'll need to build guest VMs which can have up to 4 vCPUs.

2. Last time I paid attention I was under the impression that the SMP client worked the best with 2 CPUs and required x86_64 linux.

Correct me if I am wrong with anything here. My basic idea was to run 9 VMs and see what happens. I just don't want to waste too much time with it. Set it and forget it.

[Adak]

Oh yes! I was thinking Windows server was your OS. Ubuntu x86 64 bit is a popular choice, and definitely folds SMP faster than Windows. More stable, as well.

My sticky on installing Ubuntu Linux and FAH SMP may have some links that are useful, if you want to install from a Live CD. (It has been updated, but is obviously geared to PC's running Windows, and wanting to change to 64 bit Linux for SMP folding.)

You are the guinea pig on this, I'm afraid. I don't know anyone with such a box. Nine VM's sounds great, but I don't see how you can just pick a number on such a unique box and say it's best, without some ramp up to see what's the best, and what kind of set up will also be a stable folder. Seems you have some insight into this, though. Me, I'd be checking the rig every few hours if I could. :D

Vijay has stated that the points system will be adjusted to give the true quads bigger WU's and more points than the dual's. No further details or indication of when this might happen, however.

Congratulations on this *monster* rig. I hope you'll give us a little tour later on, with some history on it, and maybe some pic's as well. :)

[Zerix01]

Each instance of FAH SMP creates four processes. On a quad most of the time not all of each core seem to be utilized. I think on average each core ends up at 75% to 85% load. So to make better use of each core it is common to run two instances of FAH on a quad so each core gets two processes, this will cause the PPD to decrease per instance but your total PPD between the two will be greater than the single instance. This also makes better use of each cores full capacity.

Using this thought process I would run 8 instances so you would end up with 32 processes running with on average (depending on work balancing and OS/ VM related activities) 2 processes per core. This will provide some great stress overall. You may want to try 12 instances (48 processes, or 3 per core) which may or may not yield a higher PPD, I really can not predict how a system like that will scale. Just remember not all WU's are are created equally and not all processes will require the same amount of CPU power all the time for the duration of the WU.

I would suggest start with 8 instances since that is fairly predictable as to how the system sill handle that. Then I would start adding instances one by one. Use FAHMON to monitor how long each WU is taking and make sure they will be sent before their due date. Once they are getting close to the due date, stop adding instances. You may not get better PPD by doing this (although you may to a point, please monitor this), this will provide a very large amount of stress and show how well the system multitasks under such stress.

Now depending on the type of load and multitasking you wish to simulate I can't really suggest how many instances per VM you should use assuming ESX allows you to select how many processors you wish the VM to see. If ESX is limited to two CPU's like VMware player is, then only use one instance per VM.

And one last note, I would run the Linux client. More efficient overall and more stable.

[{PMS}fishy]

So far I have built 8 VMs, configured 3 of them, and in the process of making more.

[{PMS}fishy]

Use FAHMON to monitor how long each WU is taking and make sure they will be sent before their due date.

Can that monitor all the machines over the network, or do I have to install it on each machine?

[Zerix01]

Can that monitor all the machines over the network, or do I have to install it on each machine?

Yes one instance can monitor all the clients as long as FAHMon has access to them. It doesn't matter the method just as long as they are mounted somewhere it has access to.

[{PMS}fishy]

Yes one instance can monitor all the clients as long as FAHMon has access to them. It doesn't matter the method just as long as they are mounted somewhere it has access to.

Any info on that? I clicked around, and didn't see much.

[dfonda]

Someone has something around on a quick way to clone your VM's...but if it wasn't Adak...I can't remember who posted it. I think they said it takes just 20 minutes or so to clone off a whole bunch.

Thats if you haven't finished already.

[cw823]

Any info on that? I clicked around, and didn't see much.

I normally will map drives to VMs, even if they are on another PC. You can just point fahmon to the network share but if it takes other login credentials and you reboot, it won't talk to them, that's why I like mapping drives.

Someone has something around on a quick way to clone your VM's...but if it wasn't Adak...I can't remember who posted it. I think they said it takes just 20 minutes or so to clone off a whole bunch.

Thats if you haven't finished already.

Actually, I have just taken the folder I setup a VM in, let's say C:\FAH\VM1, copied it, renamed it, booted it, renamed the machine in ubuntu, and it folds away. Maybe there's a longer way, but that has worked for me.

[{PMS}fishy]

I normally will map drives to VMs, even if they are on another PC. You can just point fahmon to the network share but if it takes other login credentials and you reboot, it won't talk to them, that's why I like mapping drives.

Actually, I have just taken the folder I setup a VM in, let's say C:\FAH\VM1, copied it, renamed it, booted it, renamed the machine in ubuntu, and it folds away. Maybe there's a longer way, but that has worked for me.

These are all Linux machines, and I can't setup NFS exports on them. Not to mention that Windows doesn't do so well with NFS. Does the program need access to the "live" files, or will it do with a 5 min copy? I can setup a cron job to dump the files from the VMs somewhere.

I'm all set with making the VMs. I'm at 16 right now, and will see how it runs before I change anything else.

[cw823]

These are all Linux machines, and I can't setup NFS exports on them. Not to mention that Windows doesn't do so well with NFS. Does the program need access to the "live" files, or will it do with a 5 min copy? I can setup a cron job to dump the files from the VMs somewhere.

I'm all set with making the VMs. I'm at 16 right now, and will see how it runs before I change anything else.


Well, with ubuntu it's as simple as installing samba and setting up the file share, so this may be a bit over my head.

You can do stuff with linux other than run FAH? :D

[{PMS}fishy]

Well, with ubuntu it's as simple as installing samba and setting up the file share, so this may be a bit over my head.

You can do stuff with linux other than run FAH? :D

Yeah, no samba for me. Too many security problems.

I have other Linux machines that do things. These will only ever fold.

[Zerix01]

Yeah, no samba for me. Too many security problems.

I have other Linux machines that do things. These will only ever fold.

Well no need for Samba if it is just a self contained machine. Have one partition setup that stores all FAH instances (in separate directories) have each VM mount that partition during boot. Then if you are using a script to auto start the clients then point the script to the new location. fahmon only needs to see that one partition to work.

I think you can do this a few ways, but just give up some space n the host OS that the VM's can all access.

[{PMS}fishy]

Well no need for Samba if it is just a self contained machine. Have one partition setup that stores all FAH instances (in separate directories) have each VM mount that partition during boot. Then if you are using a script to auto start the clients then point the script to the new location. fahmon only needs to see that one partition to work.

I think you can do this a few ways, but just give up some space n the host OS that the VM's can all access.

There are 2 hosts. The host OS is not Windows.

[Zerix01]

Well if you are not willing to setup samba on any of the VM's then you can just run two copies of FAHmon, one on each system. Tally up the total PPD reported from both, I think this would still be an easier to track what is going on.

FAHmon is cross platform. Everything else I described was with Linux from VM's to host OS in mind. If you have questions on how to do any of that then let me know.