On another note: can we run BOINC in a VM under ESXi? Is it worth it? What's the best way to setup? 1 VM per host? Multiple VM's per host?

All insight and input is mostly appreciated!

TIA

Yes, there's nothing about a VM that would make it inherently unsuitable for running CPU workloads. GPU workloads are more tricky, but that's not going to be an issue for Rosetta.

That being said, if you're thinking of using VMs as a way of running BOINC alongside other workloads on the same physical machine, be aware that modern processors will drop their clock speeds (sometimes significantly) when running a heavy workload for a long period of time. This can negatively affect the performance of other workloads running on the same physical machine.

Hello, I am not planning on running other workloads on the same machine. Since my hosts are running ESXi, there is not even a possibility to run workloads on that OS bare metal. So only VM's.

Question is: 1 VM with all vCPU's allocated or multiple VM's with only 1/some vCPU' allocated?

I would tend toward a single large VM (leaving at least one processor available for the host), as that should minimize the overheard of the guest OS.

But if you have multiple equally-powerful machines, it might be interesting to experiment with different guest sizes, and see which approach provides the highest throughput.

Caution, R@h has enough variation in credit, that is makes a poor benchmark for performance comparison, especially in short (less than one week) windows of time.

---

Rosetta Moderator: Mod.Sense

Caution, R@h has enough variation in credit, that is makes a poor benchmark for performance comparison, especially in short (less than one week) windows of time.

I do not understand what you are trying to say. Can you please elaborate and explain further?

William Albert suggested experimenting with various host sizes, and seeing which produce the best results. I am really just trying to say that it will be very difficult to measure which is best. In fact if you look at total credit for the last 24 hours and the next 24 hours, you could easily see a 5 or 10% difference without changing the machine at all. So, using Rosetta credits as a measure of performance would be less than ideal.

---

Rosetta Moderator: Mod.Sense

William Albert suggested experimenting with various host sizes, and seeing which produce the best results. I am really just trying to say that it will be very difficult to measure which is best. In fact if you look at total credit for the last 24 hours and the next 24 hours, you could easily see a 5 or 10% difference without changing the machine at all. So, using Rosetta credits as a measure of performance would be less than ideal.

Thank you. For the moment, I will set up 1 VM per physical host and allocate as much vCPU and vMEM as possible. Until somebody here tells me otherwise. ;)