HyperVisor
-
@DustinB3403 said in HyperVisor:
@Dashrender said in HyperVisor:
What year did the Dell R300 come out? the R710 was from 2009 - thats ancient from a hypervisor POV. I'm sure moving up as you mention have that in mind, but that's not all, the performance increase is still likely to be noticeable.
But at what cost? I understand the value gain for sure, but this is a topic posted by @mroth911, who's topics border on the clinically insane side of things (again no offense mroth).
As a PoC to just setup and see where and how things should be setup, using these in a lab scenario is perfectly acceptable.
But PoCing old tech - assuming he can't run the latest versions of Hypervisors - is worthless. Why would you PoC old stuff? God only hopes you aren't deploying that old stuff.
wasn't @mroth911 talking about putting in all SSD storage? That will likely cost a lot more than the server - and he's then likely to run into bottle necks in the server instead of the storage.
-
@Pete-S said in HyperVisor:
@notverypunny said in HyperVisor:
@Pete-S said in HyperVisor:
@Dashrender said in HyperVisor:
Assuming Moore's law holds, 7 years newer equipment, you're looking at 6 to 8 times faster gear.
It doesn't hold unfortunately.
Per core it's 15-20% faster per generation when there is a major technology shift. A lot less otherwise.
So R710 is Nehalem Xeons.
We have the following major generations:
- R710 - Nehalem architecture, Xeon 5500 series, on 45 nm
- R720 - Sandy Bridge architecture, Xeon E5-2600 v1, on 32nm - PCIe 3.0 introduced on E5-2600 v2 series.
- R730 - Haswell architecture, Xeon E5-2600 v3, on 22nm - DDR4 RAM introduced
- R740 - Skylake architecture, Xeon Scalable, on 14nm
Expect cores on a R740 to be roughly 70% faster than R710 at the same GHz. It's a lot but not as much as you would think. Especially since clock speeds have gone down and core count has gone up.
We've got a mix of 720 and 730 units in production and the 730s deal with the spectre / meltdown garbage much better than the 720s.
That's interesting to hear. Did you notice this on the production workloads or is it from running benchmarks?
Production workloads, various versions of XenServer running W10 VDI instances.
-
At some point old tech just becomes too old.
R710 for instance doesn't have pcie 3.0, doesn't have sata 3 on the cpu, doesn't have aes-ni instructions (big deal for encryption) and it's power hungry.
I suggesting R720 with Ivy Bridge CPUs, E5-2600 v2, as the oldest tech that is still decent today.
Then you have 22nm technology, 12 core CPUs, pcie 3.0 (means you have the I/O capacity for fast raid cards, pci based flash etc). You have sata3 on the cpu so you can use sata ssd to it's full potential. You're also on the last generation of CPUs that uses ddr3 RAM so you can load up the server with lots of GB for a modest price.
-
@Pete-S said in HyperVisor:
At some point old tech just becomes too old.
R710 for instance doesn't have pcie 3.0, doesn't have sata 3 on the cpu, doesn't have aes-ni instructions (big deal for encryption) and it's power hungry.
I suggesting R720 with Ivy Bridge CPUs, E5-2600 v2, as the oldest tech that is still decent today.
Then you have 22nm technology, 12 core CPUs, pcie 3.0 (means you have the I/O capacity for fast raid cards, pci based flash etc). You have sata3 on the cpu so you can use sata ssd to it's full potential. You're also on the last generation of CPUs that uses ddr3 RAM so you can load up the server with lots of GB for a modest price.
The performance jump from DDR3 to DDR4 might make it worth it for the 730. Depends on your use-case / workloads.
-
@Jimmy9008 said in HyperVisor:
Interesting video. Thanks for that. Where do you get the rough number of nines figures from for various kit?
Just rough calcs of what true HA is likely to do when you do "almost certain success of failover" combined with the failure rates assuming that their fail events are unrelated.
It's all a little grey because there are loose numbers in there like the node recovery time. But say it is 6 hours on average. Take six nines and multiple that by like 1500, you start getting eight or nine nines.
-
@notverypunny said in HyperVisor:
@Pete-S said in HyperVisor:
@Dashrender said in HyperVisor:
Assuming Moore's law holds, 7 years newer equipment, you're looking at 6 to 8 times faster gear.
It doesn't hold unfortunately.
Per core it's 15-20% faster per generation when there is a major technology shift. A lot less otherwise.
So R710 is Nehalem Xeons.
We have the following major generations:
- R710 - Nehalem architecture, Xeon 5500 series, on 45 nm
- R720 - Sandy Bridge architecture, Xeon E5-2600 v1, on 32nm - PCIe 3.0 introduced on E5-2600 v2 series.
- R730 - Haswell architecture, Xeon E5-2600 v3, on 22nm - DDR4 RAM introduced
- R740 - Skylake architecture, Xeon Scalable, on 14nm
Expect cores on a R740 to be roughly 70% faster than R710 at the same GHz. It's a lot but not as much as you would think. Especially since clock speeds have gone down and core count has gone up.
We've got a mix of 720 and 730 units in production and the 730s deal with the spectre / meltdown garbage much better than the 720s.
I've heard from multiple places that there is a solid leap in the 3 series procs over the older ones.
-
this is my current scale setup.
-
@mroth911 said in HyperVisor:
this is my current scale setup.
Is your goal to remove your Scale solution and implement something built in house?
I feel like we've had this conversation here on ML in the past and you were wanting to setup a HA fleet for production rented VM space.
-
No I don't want to replace scale. I just want to build something for testing. And or use if for my own in house stuff. The production is staying on scale.
-
@mroth911 said in HyperVisor:
I just want to build something for testing
Okay, thank you for the clarification. To ask a followup, is a lab worth spending a few additional thousand on for SSD arrays? Or spending even more for more modern equipment?
I personally would just use what I had unless there was an obvious reason that I had to spend the money.
-
@DustinB3403 said in HyperVisor:
@mroth911 said in HyperVisor:
I just want to build something for testing
Okay, thank you for the clarification. To ask a followup, is a lab worth spending a few additional thousand on for SSD arrays? Or spending even more for more modern equipment?
I personally would just use what I had unless there was an obvious reason that I had to spend the money.
Alone these lines - I too am trying to understand the need for a HA or semi HA setup for a lab? That's a lot of build out for something that's just lab.
If the idea is to mirror production - well, this won't qualify for that either - you'd need another Scale so you can compare mostly apples to apples.
-
@DustinB3403 true.. however I am trying to learn the technology . The problem I am have is trying to learn. Thats how Scale and all these other companies come in business. they take an idea and make it better.
I know everyone has they're opinions about money and what not. But at the same time... this is what fuels my desire in the tech industry. Is to learn new tech and keep evolving over time. Seeing what other best practices are and so forth and so on.
So it might look at a costly investment etc, and a lot of time wasted but that is your opinion. Its just like collection cars. this not only how I make money but a hobby for me too.
