Effective Testing for Multiprocessor Systems

[davecason]

I had an idea and I wanted to throw it out to the group to get feedback.

VMWare separates virtual machine processors onto separate physical or hyperthreaded processors when more than one virtual machine is running (http://www.vmware.com/support/ws5/doc/intro_vmspec_ws.html). Wouldn't it make sense to use a pair (or more) of virtual machines to help test the capabilities of multiprocessor systems (HT, physical, dual core, or a combination)? By using a test scenario that somehow incorporated virtual machines, you may be able to show a clearer picture of how well a system may actually perform under the stress of excessive multitasking. A test performed by a single computer running multiple virtual servers and workstations seems like a great way to really stress out the new hardware entering the market this year.

I became curious about this while trying to determine which hardware scenario will best accommodate my situation at home. I have been slowly replacing all of the extra computers lying around my house with a single computer (with lots of RAM) running virtual machines. This computer isn't currently on a multiprocessor system but I will convert it to one by end of year.

My best case scenario appears to be Tyan?s Thunder K8WE nForce4 Pro with the 2200 and the 2050 chips utilizing a pair of single or dual core opterons and memory dedicated to each processor. I wonder if the memory bandwidth advantages of this chipset are outweighed by lack of processor speed available to a dual-core, dual Hyperthreaded chip from Intel running at a higher clock speed.

Please post your thoughts (and corrections).

Thank you!

 

[uOpt]

That's not a good idea as the hardware and OS use of a virtual machine doesn't resemble anything a normal application would.

In particular you probably have a system call and virtual memory mapping "storm" going on which will make most SMP-capable OS kernels lock out other applications for a much bigger part of the time than normal applications.

 

[davecason]

I am not looking for a way to test normal computer usage. Most people use their computers at home for a single task, such as playing a game or reading on the Internet.

VMWare Workstation 5 is supposed to have better memory management features, similar to those found in their high-end ESX server software. I also suspect that modern OS's such as Windows XP SP2, 2003 SP1, and any version of Linux after 2.2 kernel will not suffer drastically from any sort of VM swapping storm. If they do, this may be a good way to expose the potential for these problems before choosing an OS for a high capacity web server or database server.
Read for yourself:
http://www.surriel.com/lectures/linux24-vm.html

Since my intention is a virtual mainframe to replace about ten or so actual computers, I am looking to actually do this so it will be "normal" to me. For the 3 dozen application delivery engineers where I work, the use of multiple virtual machines for application testing is already "normal" for them. Perhaps as the paradigm shifts, we should try to get ahead of the game and find out which hardware platforms will best handle this kind of software stress.

As new hardware arrives, such as the Tyan K8WE, that is capable of supporting 16GB of RAM at a somewhat reasonable cost, I am beginning to wonder how practical virtual memory (VM) will continue to be going forward. At some point, I suspect that VM will become a thing of the past or reserved for very specific scenarios where an enormous amount of physical RAM is somehow exhausted.

I appreciate your 2 cents but I?m a little curious why you were so quick to shoot down the idea?

 

[bsobel]

Since my intention is a virtual mainframe to replace about ten or so actual computers, I am looking to actually do this so it will be "normal" to me.

I've been doing this under GSX for years now. Try to have as many utility servers/workstations as VM's running off the physical home server. Sure makes new machine migration nice...

I am beginning to wonder how practical virtual memory (VM) will continue to be going forward. At some point, I suspect that VM will become a thing of the past or reserved for very specific scenarios where an enormous amount of physical RAM is somehow exhausted.

Not going to happen. Way too much benefit of having a large and unique address space per process that you can map back to physical memory on demmand.

I appreciate your 2 cents but I?m a little curious why you were so quick to shoot down the idea?

Maybe you didn't explain the idea clearly enough (I read your post a number of times). But Vmware, Xen, etc are great for what your using them for. As a host stress tool, not as well. Vmware does some pretty weird things during the world switch that is just different from how most things work, it's not stressing the system in the same way as other applications do.

Bill


[/quote]

 

[davecason]

Thanks, Bill.

Vmware does some pretty weird things during the world switch...

I agree, but the new versions of the software appear to handle things a little differently. Let me know if you've tried any of the new software. I also wonder if this will continue to matter when the software solutions running on top of the OS are replaced by solutions closer to the hardware:
http://www.amd.com/us-en/Corporate/VirtualPressRoom/0,,51_104_543~96162,00.html

Not going to happen. Way too much benefit of having a large and unique address space per process that you can map back to physical memory on demmand.

This may be true for 32-bit platforms but it becomes much less important in the 64-bit space. If you plan to exceed 4GB of RAM on a 32-bit platform there is no way to get around it, but we've got a little more wiggle-room in the 64-bit world. I hope to give it a shot by year's end. Has anyone explored this with a virtual machine solution?

Maybe you didn't explain the idea clearly enough (I read your post a number of times)

I apologize. I don't get a lot of free time for this kind of thing. I was in a hurry.

 

[bsobel]

I agree, but the new versions of the software appear to handle things a little differently. Let me know if you've tried any of the new software. I also wonder if this will continue to matter when the software solutions running on top of the OS are replaced by solutions closer to the hardware:

Not much differently (I know the Vmware stuff extremely well, spent alot of time with Mendell and the gang over there before the EMC deal). As for will this continue to matter, no, things are changing quickly with hardware support. Xen should start giving them some real competition (at least on the ESX side) moving forward.

This may be true for 32-bit platforms but it becomes much less important in the 64-bit space.

It's not less important, we have lots of applications that will use that address space. Thinking differently is much like Gate's famous 640k quote.

If you plan to exceed 4GB of RAM on a 32-bit platform there is no way to get around it, but we've got a little more wiggle-room in the 64-bit world.

Sure there is, the solutions are ugly, but their are solutions.

Cheers,
Bill