Ghost VM Instance?

[capreppy]

Q6600 @ 3.2
Running GPU2 Client on eVGA 8800GTX (~5200 PpD)
Running Linux SMP Client via VMWare Player 2.04. Instance was created on lappy using VMWare Server 1.06

I upgraded the GPU client yesterday to 6.20. Ran it by itself and things were ok. I fired up the Linux SMP instances via VMWare Player. Let it run for about 30 minutes. Checked FaHMon and well, the results were kinda pissing me off.
GPU2: 550 PpD
SMP1: 3500 PpD
SMP2: 3400 PpD
Although the SMP clients are set to Idle, they were taking all available CPU cycles (at least one of them and the other was at 97%).

I killed one of the instances of the Linux SMP Client via VMWare Player. I confirmed that the instance was gone via task manager. So at this point, 1 SMP Client and 1 GPU2 client. Let it run for the a couple of hours while my wife and I watched a movie together. After the movie (note: Do NOT rent Goya's Ghost, it sucked BIG TIME), I went and checked FaHMon again.

This is where it gets TOTALLY WEIRD!!! FaHMon is showing the GPU2 client averaging 5200 to 5300 PpD (good, this is what was expected). The SMP client that is still up and running is there and is getting 3000 PpD. The second SMP client that I killed is apparently running. I went to the box and checked task manager. 49% to 50% utilization for the SMP that I can see. 2 to 5% for the GPU client. NOTHING ELSE!!!

Somewhere or another, this second Linux VM SMP is being processed. It's only getting 1800 PpD, but that's pretty good considering I can't find it. I've checked and double checked task manager. I've checked the folder that the VM Instance resides. There are NO locks on the VM so it isn't supposed to be running.

I am ABSOLUTELY at a loss here. The box only has VMWare Player. It does not have VMWare Server installed at all.

Any thoughts?

 

[nomad8u]

Before I started running dual GPU I ran 2x notfred SMP clients (using server not player) with the GPU client. Got normal point on the GPU and 1500-1800 per SMP.

When starting my clients with that setup I always started the GPU first and locked it to core 3. Then started both SMP VMs (one at a time and letting them both initialize) and after they got up, using SMP Seesaw Pro to move both "vmware.vmx" files to core 0, 1 and 2. Not sure WHAT the hell you're seeing there... It does rather look like both VMs are running. Seems to be putting out some nice points that way though..

You could try using Process Explorer to look for it. Great little program that I use in place of task mangler.

 

[BillR]

What you are seeing has happened before when VM was the hot setup. I suggest a double reboot and for now your ghost should go away.

Luck

 

[SmokeRngs]

I'm just curious how you were getting 3400 and 3500 PPD on the VMs if they were both running at the same time. I have my main system (in the sig) with Vista64 running dual VMs and the GPU client and I'm only pulling about 2200PPD out of each VM. I have another Q6600@3.6 with openSUSE running dual VMs of openSUSE and it does around 2300-2400PPD on each VM.

 

[capreppy]

I think FaHMon was giving me bad numbers. Now that I think about it, those numbers were obviously way off

My wife is using this box and when she's working, she shuts down all of the SMP Clients (she doesn't like all of that stuff cluttering her task bar and there is nothing I can do about it).

She shut down the visible SMP client when she started working this afternoon and the Ghost VM is still running at ~1600 PpD. I'll reboot tonight after she's done working and see what happens. It should obviously kill the ghost VM.

The reason I killed the second VM is that the two Linux VMs were choking the GPU2 client. Both are set to idle, but still pull nearly 98% total CPU. The GPU2 client in Vista doesn't need that many CPU cycles, but 2% isn't enough.

The reason I'm using the Linux VMs as opposed to notfred's (which I did try and posted my thoughts on it) is the fact that you can't use the -advmethods flag to get the A2 core. The A2 seems to be more robust than the A1 core (at least these are my observations). But it seems that on either one, you can't set the CPU usage.

Is there a way to limit the Linux VM to say 95% of two CPUs? If I could limit the Linux VMs to 95% each, that should give enough cycles for the GPU2 client to not go hungry.

 

[SmokeRngs]

Since you're running Vista, there shouldn't be any trouble running two VMs and the nVidia GPU client.

What you'll need to do every time you reboot is in Task Manager, set the two VM instances to idle priority. By default they are set to normal priority instead which makes the machine a little sluggish as well as keep away the needed CPU cycles from the GPU client. I run the VM instances on my machine at idle priority and the GPU client at normal priority. So far this has worked out well for me.

 

[APOLLO]

What you'll need to do every time you reboot is in Task Manager, set the two VM instances to idle priority. By default they are set to normal priority instead which makes the machine a little sluggish as well as keep away the needed CPU cycles from the GPU client. I run the VM instances on my machine at idle priority and the GPU client at normal priority. So far this has worked out well for me.

How does that affect the clients in the VMs? Any perceivable slowdowns by changing priority?

 

[SmokeRngs]

How does that affect the clients in the VMs? Any perceivable slowdowns by changing priority?

There is a small hit. I might lose 100-150PPD per VM but it's difficult to tell overall because this is my main machine and I use it for a lot of other things. My Q6600@3.6 which does nothing but fold does 2250-2300PPD per VM instance and it's running openSUSE 10.3 with two openSUSE 10.3 VM instances. Well, it might be two openSUSE 10.2 instances. I've had the VMs around for a while and have been switched from different machines.

Either way, I get around 9600PPD total out of the machine if it's left sitting and that's with an 8800GT with shaders at 1800. However, I have to say that after upgrading to the newer GPU console only client (as the systray client would crash during the install) that I've been able to get just a bit more PPD from the GPU client with no other changes. It will top out around 5400PPD (according to FAHMon) whereas before it would never top out above 5120PPD. However, it will only do that when the machine has been totally idle with no use whatsoever and usually stays closer to 5200PPD. The 5200PPD figure for the GPU client and the 2200PPD figure for each VM is what I used to figure up 9600PPD for my total PPD for the machine.

For the most part, changing the priority to idle from normal on the VMs does not seem to affect the PPD very much. However, I setup the machine this way from the beginning and did not leave the VMs at normal priority once they were installed and started running. The only real comparison I have is with the system running the same speed (only half the RAM at 2 gig but should not make a difference because all VMs are set to use 512meg) but the second system is running openSUSE 10.3.

System stats:

Vista64
Q6600@3.6
Gigabyte P35-DS3R
4 Gig of RAM
2 openSUSE 10.3 VMs
1 GPU client 8800GT with shaders at 1800

openSUSE 10.3
Q6600@3.6
Abit IP35-e
2 Gig of RAM
2 openSUSE 10.3 (or 10.2) VMs
1 console client in the host at idle priority to suck up stray cycles.

I hope that information helps a bit.

 

[APOLLO]

Thanks for the info. I have only one VM running in my main machine so far. It has become sluggish ever since I had it installed. I'll try to lower the priority level and see if it makes a difference in responsiveness. There's two settings I could try, 'below normal' and 'low.' Thanks again.

 

[SmokeRngs]

Thanks for the info. I have only one VM running in my main machine so far. It has become sluggish ever since I had it installed. I'll try to lower the priority level and see if it makes a difference in responsiveness. There's two settings I could try, 'below normal' and 'low.' Thanks again.

Even with two VMs running I don't notice any sluggishness at all. Normally, the sluggishness is related to the VMs running at normal priority which means it's very difficult to get CPU cycles away from the VMs when you're running something in the host. Normally, this isn't a problem if VMs are used as they were designed, mainly as a way to make one box take over the duty of a few smaller boxen. Basically, an enterprise environment.

I don't know if this matters, but I am using VMWare Server on all my systems. I don't create the VMs on one machine and move them to another machine. I create each VM instance on each machine as I set it up. I know it's not the most efficient way to do things but I also don't have the problem with the clocks running fast or slow in the VMs that I create for each machine. The exception are the VMs I moved from one machine to another. The clocks on the VMs I moved have nothing to do with real time which is very annoying when trying to monitor things with Fahmon since half the day Fahmon says the clients are hung when they really aren't. I do have those two VMs setup to readjust times using the Network Time Protocol in openSUSE but currently it only does that once a day.