Virtualized OS & Multicore CPUs

[Bozo]

Is it possible to assign a virtual OS to a particular core of a multicore CPU? Or is that what some motherboards are doing with the Hardware Virtualization setting?

Bozo :joker:

 

[ddrueding]

I thought the setting in the motherboard enabled certain CPU code that helped with virtualization performance. I haven't seen a setting in VMWare (Workstation or Server) that would allow you to specify a core, but in task manager you can right-click on a process and "set affinity" to a core or cores. Be careful, though, this just crashed my computer*

*and interestingly enough, after rebooting from a hard reset, Firefox 3 Beta 5 kept this partially typed post! Wow!

 

[sechs]

I'm not sure what the advantage of assigning cores would be. Actually, it seems kind of limiting, resource-wise.

In general, I'm one to be running one VM, which the operating system will generally run on one core, or multiple, which need to be able to move between cores to remain smooth. So, that might be colouring my point of view.

Why do you want to do this?

 

[Fushigi]

Its definitely possible. IBM does it on their POWER hardware (POWER5 & 6 CPUs). Assign a partition to have an exact whole number of cores and it becomes essentially dedicated. But then that partition can't take part in dynamic movement of CPU between partitions. There is a minor performance boost v. a non-dedicated core, but only by 2-4 percent. I don't know where the performance comes from; maybe the hypervisor assigns it and leaves it alone (no task swapping related to partition overhead). Maybe the core has better cache utilization. Maybe it's something else entirely.

I don't know if any of the x86 VM engines have the same capability.

 

[Bozo]

I'm not sure what the advantage of assigning cores would be. Actually, it seems kind of limiting, resource-wise.

In general, I'm one to be running one VM, which the operating system will generally run on one core, or multiple, which need to be able to move between cores to remain smooth. So, that might be colouring my point of view.

Why do you want to do this?

More curious than anything. But, with multiple core CPUs, and most applications not optimized to use more than one CPU, it would make sense to assign each VM to a core of it's own. At work we are discussing the idea of having a computer with three operating system. One host OS and two VMs. Each VM gets assigned it's own memory, has it's own hard drive, and posibly its own CPU, using a multiple core processor.
This would be more cost effective than separate computers.

Bozo :joker:

 

[ddrueding]

I regularly assign VMs their own real hard drive and a chunk of real RAM, but I don't think you will see that much speed improvement trying to out smart the CPU scheduler. Besides, one of the cost savings in VMs is the ability to assign more VM resources than you actually have, letting different servers "max out" at different times and share the CPU/RAM.

 

[Handruin]

VMWare's ESX server allows you to assign a core (or multiple cores) to a specific guest OS, though they don't look at is as cores, but physical and logical (hyperthreaded) CPU's. We don't use this feature because I felt it was better to rely on the built in SMP features they've implemented to make better use of idle CPU time.

One of the better features you can use in something like ESX is the DRS feature. When you have multiple physical servers in an ESX farm with vmotion enabled, it will monitor and migrate busier guest VM's onto physical systems that are less busy based on loads. This can be configured for fully automatic migrations. Since vmotion allows migrations with live machines, the end user doesn't notice they've been moved to have their needs fulfilled by a machine with more resources.

In regards to the Intel Virtualization Technology, the way it was explained to me when I attended a presentation at VMWorld is that it is late to the game. The engineers at vmware explained how they were able to handle the basic functionality through software more efficiently than the hardware implementation as it is built today. When using a product like ESX server, the only reason you will enable the VT in the bios is to be able to use a 64-bit OS. Everything else is still done faster through software's VMM as far as I know right now with the ESX 3.x software.

 

[sechs]

Its definitely possible. IBM does it on their POWER hardware (POWER5 & 6 CPUs). Assign a partition to have an exact whole number of cores and it becomes essentially dedicated. But then that partition can't take part in dynamic movement of CPU between partitions. There is a minor performance boost v. a non-dedicated core, but only by 2-4 percent. I don't know where the performance comes from; maybe the hypervisor assigns it and leaves it alone (no task swapping related to partition overhead). Maybe the core has better cache utilization. Maybe it's something else entirely.

Partitioning is not virtualization in the sense that we are talking about, so they don't really compare.

A partition is a subset of hardware that appears as a separate machine. The operating system actually runs on the partition hardware, whereas a "virtual machine" runs in software that runs on the hardware. In general, partitioning is done in hardware, and, for all intents and purposes, is a separate machine with the assigned resources; a "virtual machine" generally emulates the hardware.

 

[Fushigi]

They are VMs. IBM has even implemented a virtualized x86 instruction set so POWER can run x86 Linux.

 

[sechs]

Do we really have to argue semantics now?

x86 emulation (not even IBM uses the term "virtualization") is done in software.

 

[Fushigi]

You brought it up.

 

[sechs]

No, *YOU* brought it up. And you can't even defend yourself, so it'd be a pretty lop-sided argument. This is, of course, on top of the fact that you're wrong.

Let's cut to the chase: You admit defeat, and I'll gloat.

Can we move on now?

 

[Fushigi]

You said "Partitioning is not virtualization". You brought it up.

Sorry you can't admit to your mistakes. I would hate to be around you in real life.

 

[Stereodude]

Do you two need a ruler to speed up your measuring?

 

[sechs]

Fushigi brought up partitioning in post #4 of this thread. But since he can't read what he wrote, he certainly wouldn't be able to read a ruler....

 

[Stereodude]

Fushigi brought up partitioning in post #4 of this thread. But since he can't read what he wrote, he certainly wouldn't be able to read a ruler....

And since you're a castrated bleeding heart liberal, you don't have anything to measure with the ruler, so I guess I'll just keep it.

 

[sechs]

And since you're a castrated bleeding heart liberal, you don't have anything to measure with the ruler, so I guess I'll just keep it.

First, let me say that this is a really sad comment, smiley or not.

Second, I'll point out that being castrated has no bearing on penis size. In this context, your comment really says something about American education.

Finally, while I realise that the "discussions" here often get quite heated, this kind of behavior is uncalled for. This comment is totally off-topic, and the attack can't even be remotely linked to the discussion at hand.

Get a grip people. It's not that hard.

 

[Stereodude]

You seem to be quite the expert on castration there. I'll just leave you to your area of expertise.

ps, it's called a joke.