What are the tradeoffs between a VMware VM and a local machine for scientific CPU intensive computing?

I have a computing task that will probably run non-stop for several weeks with occasional restarts. My IT folks seem to feel a VMware machine set up on our corporate servers would be a better solution than purchasing a local machine (XEON 5550, 10k drive, 4gb ram, due to 32-bit software). It seems logical to me that server hardware would be better for high-traffic multi-user tasks, but what about low-traffic single user computation. My experience is solidly in the desktop world with little to no networking experience. Thanks for any help.

Additional clarifications due to comments below.

The process will be very CPU intensive and may consume 4 cores at 100% for long periods of time. This work is part research/part pilot project. One of the outcomes will measure the length of time it takes to process imagery on a per acre basis. This will be affected by both computing speed and automation of multiple disparate process steps.

From serverfault kpierce8

• From your own point of view, VM will not produce any advantage. As long as your IT folks understand the purpose of that VM and allocate CPU and memory resources to that VM equal to what you'd have with a physical server, you will see pretty much the same result.

  However, for the IT department this would be a much better solution: no new server to maintain; no expense; no space used in their racks; significantly lower additional power consumption.

  Bottom line: let them do what they're paid to do; just make sure they understand what you need.

  kpierce8 : Thanks. By local machine, I meant one at my desk that I primarily control. The cost is covered by the grant and part of the project involves estimating how fast we can get it done. Will our terrabyte of image data take three months to process or two? Questions like that.

  ynguldyn : I'm yet to see a server that an engineer/analyst/scientist installs under their desk and that never requires attention from professional IT people. Just doesn't happen. But your other points are valid. Grant money you can't always claim for other things if you save it here, so might be worth spending on the server just for the sake of accounting. And if time estimate is a part of the research, then you need to build something that would match the environment as envisioned by the project - VM may not fit. Funny how now that you've elaborated on your question, my answer is totally different.

  kpierce8 : Truthfully I wasn't sure exactly what parts of my problem were relevant. I greatly appreciate the good feedback.
  From ynguldyn

• VM Pro's:

  1. Part of your infrastructure for monitoring and backup purposes. If it dies on managed infrastructure then it gets restarted by the operations team. If it dies on a box in your office then it gets restarted when you return to the office. Same goes for backup if it needs to be backed up just imagine how long it will take to restore your 1Tb to a Desktop PC.
  2. If it doesn't perform well enough on a VM it is a straightforward task to move it to more powerful hardware.
  3. Snapshots can be taken of the VM at different times.
  4. Cost. Cheaper to run on the VM.
  5. Often there is little or no lead time in getting it set up. e.g. do not need to purchase physical equipment. if you can have the VM tomorrow or wait 2-3 weeks for purchase order/kit sign off then the VM could be quicker.

  VM Cons:

  1. Possibly/Probably runs slightly slower (depends on the workload).

  I'd run with the VM solution but get it in writing that if it doesn't perform to your requirements they will move it to faster hardware or a physical solution.

  My own experience is moving legacy systems into VMWare so they generally see a speed boost because they are coming off +5 year old hardware onto the latest kit.

  From davey

• Personally, I'd go for the VMware cluster for a variety of reasons.

  • The amount of control you'll have over a VM will be the same as a local machine.

  • I'd think that they could get you the same amount of CPU and RAM in your VMware infrastructure as you could on a local computer. Double-check that with the IT people.

  • Odds are good that the VMware infrastructure has access to higher speed and larger capacity storage than a local computer. They can probably also run backups on your data pretty easily from the VMware cluster, too.

  • The power is probably a lot more stable in the VMware cluster than at your desk. If your computing jobs would be disrupted by a power outage then the data center is probably a better place for the job to run.

  • Finally, it will be a lot quieter to use the VMware machine versus sitting beside a workstation class PC in your cube.

  Dave M : Excellent points. Modern servers can often handle these tasks as well or better than a Workstation and can be up and running very quickly in most cases. Quality of power is also a big concern and the datacenter will be well protected (one hopes) Also much easier to backup before patching and revert to snapshots if something goes wrong.
  From Evan Anderson

• It really depends on how much CPU time it's going to eat up. The problem for a vmware solution is that if this eats up one XEON 5550 CPU (potentially 4 physical and 4 virtual cores worth of CPU time) it's fairly expensive to dedicate 1 full CPU. Saving from Vmware comes from sharing the load since it's assumed that the system that was virtualized is not going to be using the CPU 100% of the time. If IT is willing to buy a better CPU than the 5500 to cover the vmware overhead then sure. Otherwise I'd take the money they would spend on vmware licenses and you should easily be able to buy better/faster procs (take a look at the intel xeon 5492, it's older but has a faster clock speed -about .6 ghz faster base processor price goes up by a whopping 535 dollars)

  If they already have excesss CPU time to cover your needs then by all means take it, but most environments I work with there isn't 21+Ghz left over to spare (again presuming your application will fully utilize the processor you specified with a full 8 threads). It's certainly a different story if this is a singlethreaded application that would only take up 2.6ghz at max.

  Another factor to consider is vendor support. If your software vendor will not support you in a virtual environment that should factor in to your decision.

  kpierce8 : I figure I can take over most of 4 cores most of the time during primary processing. That's why I thought a dedicated machine would work best. All good points.

  Helvick : I'd echo all of the pro-VM comments elsewhere but this clarification argues against going the VM route - if the app is going to drive 4 cores at 100% for an extended period of time it is going to really stretch a VMware host. It is certainly possible but multi-core scaleout in virtualized environments is not as effective as it will be on dedicated physical kit. I'd still recommend the VM route but this would make me (as a VM infrastructure guy) think quite carefully about it.

  Jim B : @helvick I wasn't going to answer at all until a reread the question and realized it said "computing task that will probably run non-stop" To me that meant high CPU utilization which is not usually wanted in a vm environment.
  From Jim B

• Given that the machine is paid for by a grant, and that the time to process impacts your research, go with a dedicated machine. You don't want web surfing to impact your results, or vice versa.

  kpierce8 : That was my initial thought. (still is I guess, even though learning the VM environment sounds interesting).

  ynguldyn : @LuckyLindy: Most likely, you're thinking about VMs that you can run on your own desktop, in Parallels or VMWare Workstation. The discussion here, on the other hand, is about enterprise VM environments - completely different beast.

  Jess : Good point, but for something that impacts his research outcome I'd err on the side of caution. Full control over a local workstation would be preferable (to me) compared to sharing resources on a corporate server.

  Joe Internet : Not to mention that IT is _probably not_ planning to put a app that consumes 100% CPU processing 1TB of imagery on their mainline production servers, which means it gets relegated to an older, less capable machine to begin with.

  kpierce8 : IT has decided that a dedicated box would probably be better.
  From Jess