How many cores for each VM are best?

APOLLO

[H]ard|DCer of the Month - March 2009
Joined
Sep 17, 2000
Messages
9,089
Those running multiple VMs on one system, how many cores have you discovered are optimal for running the SMP client? In the past, I was forced to run 2 cores on each like everyone else, but things are different now with the release of VM Player. 3.0. Assuming memory resources are not a major concern, is there any performance reason in terms of aggregate speed (system PPD) to configure VMs with 4 cores each instead of 2? I have a lot of 2-core VMs and won't go through the hassle of installing a new version of VMware if the performance increases are negligible moving to 4 cores. Thanks for any help.
 
I currently run a four-core VM and I get better performance from it than I used to with two 2-core VMs. My PC is also more responsive and I have more RAM available due to the lower overhead.
 
My understanding is ...........

If you get A1 work-units then 2 cores per VM are still best by far.
If you want to run -bigadv work-units then you need to run 8 cores per VM.
Otherwise it makes no real difference in terms of PpD if you run 2, 4 or 8 cores per VM if you have the memory to run them all in the ram.

But I may be wrong.

Luck ............ :D
 
I run a 4 core VM as well. I am running it on a Q9550, thats a total of 2 VM's that I run on that single Q9550.

My two E8600's run a single VM each.
 
I run a 4 core VM as well. I am running it on a Q9550, thats a total of 2 VM's that I run on that single Q9550.
How are you running a 4-core VM along with another VM on a quad? You're probably losing performance from that.
 
I run a 4 core VM as well. I am running it on a Q9550, thats a total of 2 VM's that I run on that single Q9550.


How are you running a 4-core VM along with another VM on a quad? You're probably losing performance from that.

I'm guessin he meant 2 core VMs? Yeah, that doesn't make sense. :confused:

I run a 4 core VM on my i7. It gets the same PPD as running 2x 2 core VMs cuz thats how I originally configured it. So for me a 4 core VM is much easier and more efficient (memory wise). I would make it a 8 core VM if my i7 stayed cool enough @ 3.36 GHz :(.
 
I currently run a four-core VM and I get better performance from it than I used to with two 2-core VMs. My PC is also more responsive and I have more RAM available due to the lower overhead.
This is what I wanted to hear. I know for certain that systems will run much smoother because of freed RAM, however, these machines are dedicated folders and I'll never be on them except for maintenance. Zero, could you give me a rough estimate what your dual VMs were producing earlier vs your single VM is producing presently? I just want to get an approximate gauge on what possible increases I could be seeing to determine if I should go through the trouble on all my machines.

If you get A1 work-units then 2 cores per VM are still best by far.

If you want to run -bigadv work-units then you need to run 8 cores per VM.
No, this isn't -bigadv I'm inquiring about, it's standard SMP WUs. Stanford is sending out almost exclusively A2 WUs these days, it matters little to me if I see the occasional A1 WU. And if it does happen, I'd gladly see these process faster with 4 cores just to be rid of them; they take so long to complete with only 2 cores, especially the P510x.

Otherwise it makes no real difference in terms of PpD if you run 2, 4 or 8 cores per VM if you have the memory to run them all in the ram.
So in essence, your advice is to leave things as they are because I'll see negligible difference, and it's just a hassle to uninstall VMWare server in order to install Player 3.0? I did that on one machine with only four cores, and it took much longer than expected after all was said and done. That's the main reason I'm curious to know.
______________________________________________________________________

Also on my mind and related to the topic of this thread: what is the maximum number of cores that can be assigned to process regular SMP WUs? Moreover, can I configure VMWare Player to utilize atypical number of cores such as 6 cores or odd number of cores??
 
Fixing what I stated.

I have a 4 core CPU, I run 2 VM's on this CPU (Q9550).

There, sorry for the mixup.
 
Zero, could you give me a rough estimate what your dual VMs were producing earlier vs your single VM is producing presently? I just want to get an approximate gauge on what possible increases I could be seeing to determine if I should go through the trouble on all my machines.
I was seeing about 2000-2200PPD per VM after the 2.10 A2 core update. With a single VM, I get about 4500-5000PPD depending on how much load the PC is under from other things.
Also on my mind and related to the topic of this thread: what is the maximum number of cores that can be assigned to process regular SMP WUs? Moreover, can I configure VMWare Player to utilize atypical number of cores such as 6 cores or odd number of cores??
I think it's 8.
Moreover, can I configure VMWare Player to utilize atypical number of cores such as 6 cores or odd number of cores??
You can. I'm not sure what the limitations on this are, but I've tried 3 cores and it worked fine, so I suspect you can set an arbitrary number of cores (within reasonable limits). I don't know how well F@H would cope with that however.
 
I was seeing about 2000-2200PPD per VM after the 2.10 A2 core update. With a single VM, I get about 4500-5000PPD depending on how much load the PC is under from other things.

I think it's 8.

You can. I'm not sure what the limitations on this are, but I've tried 3 cores and it worked fine, so I suspect you can set an arbitrary number of cores (within reasonable limits). I don't know how well F@H would cope with that however.
Thanks for the detailed reply, I really appreciate it. For a slower 8-core+GPU box, would you recommend 2 VMs or one VM with 6 or 7 cores allocated? It's these kinds of situations where it doesn't seem like there's a clear answer to what's the best configuration now. :confused:
 
I would recommend that you use as few VMs as possible with as many cores as possible. If you're running an nVidia GPU, you shouldn't need to leave a core free for it. Just run a VM across all 8 cores.
 
I would recommend that you use as few VMs as possible with as many cores as possible. If you're running an nVidia GPU, you shouldn't need to leave a core free for it. Just run a VM across all 8 cores.
Then I'd have to set affinity higher for the GPU clients, right? Wouldn't it slow down the GPU clients otherwise if the VMs are accessing the same cores as the GPU clients?
 
You might have to set priority for the GPU clients, not affinity. Either way, that's a fairly easy thing to do.
 
On my triple core AMd CPU I run a notfred VM set to 1 instance per 4 CPUs. I have the VM configured to access all 3 cores. It runs marginally faster than it did before with a 2-core VM. This CPu is also supporting 2 ATI GPU clients, which even after tweaking the settings still use a hefty chunk of CPU to run.
 
You might have to set priority for the GPU clients, not affinity. Either way, that's a fairly easy thing to do.
Yes, I meant priority, sorry my mistake. I'm using XP on these systems, Task Manager displays full utilization of a single core and some people have stated in the past that is a misrepresentation of actual core usage. It's one of the things concerning me about having my VMs access every core, even on a 8 physical core system. What is the most highly regarded affinity utility these days excluding WinAFC?
 
Run one. Even if you have enough RAM for 4 VMs, it's just so much less hassle configuring and keeping track of just 1 VM than 4. And if you don't have enough RAM, your system will slow down and grind to a halt.

Moreover, can I configure VMWare Player to utilize atypical number of cores such as 6 cores or odd number of cores??
You can. In the *.vmx file there's a line telling the VM how many virtual CPUs to assign. Just change it to whatever you want.

Yes, I meant priority, sorry my mistake. I'm using XP on these systems, Task Manager displays full utilization of a single core and some people have stated in the past that is a misrepresentation of actual core usage. It's one of the things concerning me about having my VMs access every core, even on a 8 physical core system. What is the most highly regarded affinity utility these days excluding WinAFC?
I used process explorer for a while, it's a replacement for the Task Manager and it provides a lot more information.

I think it's 8.
I'm pretty sure it scales above 8.
 
Yes, I meant priority, sorry my mistake. I'm using XP on these systems, Task Manager displays full utilization of a single core and some people have stated in the past that is a misrepresentation of actual core usage. It's one of the things concerning me about having my VMs access every core, even on a 8 physical core system. What is the most highly regarded affinity utility these days excluding WinAFC?
I don't know what the most well-regarded program is, but I personally use Bill2's Process Manager and it works great.
 
I've just swopped my Skulltrail's box from 4x 2 core's to 1x 8 core's so I can run -bigadv work-units.
My basic PpD has probably increased by ~1k PpD, up from ~10k to ~11k.

Luck ......... :D
 
I've just swopped my Skulltrail's box from 4x 2 core's to 1x 8 core's so I can run -bigadv work-units.
My basic PpD has probably increased by ~1k PpD, up from ~10k to ~11k.
How do you configure all your GPUs to have sufficient use of processor resources if you're running -bigadv WUs? Don't you have 6 GPU clients running on your Skulltrail system? Conventional wisdom indicates a maximum of 4 GPU clients per core for optimal performance, and I have seen even that has its limitations on my 4-GPU box with certain WUs, which is why I have recently allocated 2 cores to GPU for that particular box. So, how can you have 6 GPU clients on an 8-core system and still run -bigadv to safely make the bonus deadline? What is the configuration for this system in regards to core priority/affinity for all your clients?

In any case, I have inquired in the OP with standard SMP WUs in mind and I was not thinking of -bigadv. My systems are too underpowered to really meet the requirements. I did a rough calculation and there's actually little or no advantage for me to run these WUs since I'd likely be forced to stop my GPU clients in order to free up CPU resources. In the end, it comes out to be about the same PPD or thereabouts.
 
How do you configure all your GPUs to have sufficient use of processor resources if you're running -bigadv WUs? Don't you have 6 GPU clients running on your Skulltrail system? Conventional wisdom indicates a maximum of 4 GPU clients per core for optimal performance, and I have seen even that has its limitations on my 4-GPU box with certain WUs, which is why I have recently allocated 2 cores to GPU for that particular box. So, how can you have 6 GPU clients on an 8-core system and still run -bigadv to safely make the bonus deadline? What is the configuration for this system in regards to core priority/affinity for all your clients?

I'm only running 3x 9800GTX+ cards on my Skulltrails, not 3x 9800GX2's.
The heat output from the GX2's is to much when add to the heat thrown out by the chipset + memory.

I've got the VM set to low priority and the 3x GPU's to normal.
I lock all the GPU's to the last CPU core and just let the VM run with everything else.

I must try switching off the GPU clients for a couple of hours and see if it is slowing down the VM by a significant amount.
By even if they slow the VM by ~1 min per frame that only drops the bonus by ~1k, the 3x GPU's pull in ~18k.
I think thats a fair trade off.
~18k + ~25k off ~700 watts. It pulls the system upto ~60 PpDpWatt.

Ps. Watching DVD's on the Skulltrails slows the VM by ~45 sec per frame.

Luck ........... :D
 
I've just swopped a couple of my Q6600 boxen from 2x VM's (server) to 1x VM (player)
I've gone from 2x ~2,200 PpD to 1x 4,600 PpD.
So only around 200 PpD increase.

Downside is that I had VMWare Server to auto-start the clients on system boot, I cannot find a setting to do that with VMWare Player.

Luck ......... :D
 
I've just swopped a couple of my Q6600 boxen from 2x VM's (server) to 1x VM (player)
I've gone from 2x ~2,200 PpD to 1x 4,600 PpD.
So only around 200 PpD increase.

Downside is that I had VMWare Server to auto-start the clients on system boot, I cannot find a setting to do that with VMWare Player.

Luck ......... :D

Only thing I have found is to make a shortcut for .vmx and place it in the startup folder.
Only problem is with Player is that it runs in the foreground. Server was service oriented so that was nice.
 
Back
Top