running gimps on a virtual machine

[sixblueboxes]

I currently have the following setup:

1. Machine "growler" is an AMD Phenom 9650 Quad-Core running Ubuntu 10.04 server, and it is running mprime on all four cores. Also, this machine is a VMWare host.
2. Machine "dashow" is a linux virtual machine on "growler" with one CPU. It's running mprime.
3. Machine "gimps" is a linux virtual machine on "growler" with two CPUs. It's running mprime.

My original reasoning was that running virtual machines would increase my Ghz-Days output because I'd increase the load average on my host. But... that's not true is it? I can't get something for nothing here. Since mprime runs on unused clock cycles, and the VMS are now taking up more used clock cycles on the host, they're actually slowing down the host, and the total output would remain the same.

I'm also puzzled by the way they show up on the website -- saying that their Ghz is "50.000" while the host is "2.130".

Thoughts?

[f1pokerspeed]

1) In answer to the GHz issue - VMs have emulated CPUs, therefore the arch. and specs are just guesswork.
2) With regard to workload, I imagine your throughput would decrease significantly - more than you stand to gain - by running more instances than available threads. Provided that you have a CPU with temp. throttling, this only exacerbates the problem.

I would avoid running VMs alongside a host on any DC project, as it just is not worth it.

[cheesehead]

Originally Posted by sixblueboxes

My original reasoning was that running virtual machines would increase my Ghz-Days output because I'd increase the load average on my host. But... that's not true is it?

No, it certainly isn't!

I can't get something for nothing here.

There's a reason that virtual machines are "virtual", and not the same as "real" machines. :-) And primes are real -- you can't fool Mother Nature.

Since mprime runs on unused

(unused by higher-priority tasks, that is)

clock cycles, and the VMS are now taking up more used clock cycles

(in higher-priority tasks)

on the host, they're actually slowing down the host

(I think you mean slowing down the low-priority mprime ... the host itself is chugging along at the same speed as before, but it's splitting that speed among more tasks, so it's the lowest-priority tasks that lose real time.)

, and the total output would remain the same.

(depends on what one considers "total output" to include)