![[H]ard|Forum](/styles/hardforum/xenforo/logo_dark.png){kind=logo}
Silent.Sin
Gawd
- Joined
- Jun 23, 2003
- Messages
- 969
Background: Bought a new Dell R320 for use as an internal database server to replace an aging SQL Server 2008 setup. It runs Server 2012 as a host and two HyperV VMs for production and testing environments. Usage is expected to be relatively light, not looking for critiques on the hardware config. Specs are as follows:
CPU- Intel Xeon E5-2430 2.2GHz (6C/12T)
RAM- 24GB DDR3 1333MHz (3x8GB triple channel), VM1-8GB, VM2-8GB
Storage- 1TB for OS+VM storage, 3TB RAID1+hot spare (1TB usable) for database use
My question arose when I was evaluating the hardware after I got it. I needed to provide some basic benchmarks to present to other staff why the purchase was made and how much of an improvement it is. Being a little pressed for time I went with some of the easier to run suites I could tackle within a day before I installed anything else on the servers and sent the host to core mode: PassMark, Cinebench, and 7zip. PassMark and 7zip were mostly as expected, but Cinebench showed some weird anomalies with ST and MT results. The scaling of CB vs 7zip seems very odd also. VM1 is for test and so I allocated 4 vCPUs to it, VM2 is prod and I allocated 6. This should leave 2 threads that can be used solely by the host for whatever it needs in the rare event that both prod and test are maxed out. Not sure how Windows would schedule logical vs physical in that instance but it should be enough to keep the system chugging along.
Cinebench (ST/MT/Ratio)
Host- .97/6.50/6.70
VM1- .73/3.60/4.92
VM2- .89/5.37/6.03
7zip (ST/MT/Ratio)
Host- 2792/19314/6.92
VM1- 2712/9846/3.63
VM2- 2732/13964/5.11
I expect that the ST performance of the VMs should closely resemble the host minus some overhead and I can observe that in VM2, but what the hell is wrong with VM1? I just can't wrap my head around why ST performance should be 25% lower and the fact that it doesn't match VM2 is bizarre. I've tried disabling the NUMA features of HyperV (allows VM migration across sockets), since this is a single socket system that shouldn't matter much anyway but it doesn't make a difference. In the long run it's not that big of a deal since it's the test machine anyway, but it'd be nice to know the software/hardware is working as intended.
Anyone have much experience with HyperV 2012? We've got a couple of HyperV 2008 environments but I've never bothered to benchmark them and they are running production stuff so it'd be a pain to actually get some numbers for comparison. Is this some weird hyperthreading/logical core issue? I would think if VM1 was somehow being allocated mostly logical cores instead of physical the scores would be a lot lower and I would think 2012's scheduler is more intelligent than that which seems to be indicated by the VM2 MT CB score. Is there any way to actually specify affinity for vms inside of HyperV?
CPU- Intel Xeon E5-2430 2.2GHz (6C/12T)
RAM- 24GB DDR3 1333MHz (3x8GB triple channel), VM1-8GB, VM2-8GB
Storage- 1TB for OS+VM storage, 3TB RAID1+hot spare (1TB usable) for database use
My question arose when I was evaluating the hardware after I got it. I needed to provide some basic benchmarks to present to other staff why the purchase was made and how much of an improvement it is. Being a little pressed for time I went with some of the easier to run suites I could tackle within a day before I installed anything else on the servers and sent the host to core mode: PassMark, Cinebench, and 7zip. PassMark and 7zip were mostly as expected, but Cinebench showed some weird anomalies with ST and MT results. The scaling of CB vs 7zip seems very odd also. VM1 is for test and so I allocated 4 vCPUs to it, VM2 is prod and I allocated 6. This should leave 2 threads that can be used solely by the host for whatever it needs in the rare event that both prod and test are maxed out. Not sure how Windows would schedule logical vs physical in that instance but it should be enough to keep the system chugging along.
Cinebench (ST/MT/Ratio)
Host- .97/6.50/6.70
VM1- .73/3.60/4.92
VM2- .89/5.37/6.03
7zip (ST/MT/Ratio)
Host- 2792/19314/6.92
VM1- 2712/9846/3.63
VM2- 2732/13964/5.11
I expect that the ST performance of the VMs should closely resemble the host minus some overhead and I can observe that in VM2, but what the hell is wrong with VM1? I just can't wrap my head around why ST performance should be 25% lower and the fact that it doesn't match VM2 is bizarre. I've tried disabling the NUMA features of HyperV (allows VM migration across sockets), since this is a single socket system that shouldn't matter much anyway but it doesn't make a difference. In the long run it's not that big of a deal since it's the test machine anyway, but it'd be nice to know the software/hardware is working as intended.
Anyone have much experience with HyperV 2012? We've got a couple of HyperV 2008 environments but I've never bothered to benchmark them and they are running production stuff so it'd be a pain to actually get some numbers for comparison. Is this some weird hyperthreading/logical core issue? I would think if VM1 was somehow being allocated mostly logical cores instead of physical the scores would be a lot lower and I would think 2012's scheduler is more intelligent than that which seems to be indicated by the VM2 MT CB score. Is there any way to actually specify affinity for vms inside of HyperV?