Bad memory performance on Windows 10

[Anti-Ctrl]

I have test in WinRAR:
Proxmox: ~7600 KB/s
Bare metal: ~8900 KB/s
Looks like 17% performance hit, not 20x as AIDA64 shows. So, i think, it's AIDA64 measurements bug in VM.

Even if AIDA has a Problem there, that doesnt explain how 3 different Tests in the Phoronix Test Suite take like 10 Times longer, than in a Linux VM. Makes no sense. Proxmox has a Problem and I´m not the only one "complaining" about sluggish Performance, even with all Drivers and GPU Passthrough. But the one who made the Thread and I have (maybe) narrowed it down to a Memory Problem. At least thats what i suspect here.

EDIT: WinRAR Benchmark on Baremetal (same Platform but only one CPU) ~14500KB/s. On Proxmox ~3500 KB/s. Something is wrong.

 

[Gektor]

Hyper-V guest VM (2019):


Bare Metal:


So, if you need near (or even - better) bare metal results in AIDA64 benchmarks - you need Hyper-V host server )))

 

[Gektor]

I can agree, that Proxmox is very good (even - ideal) for UNIX-like systems virtualization. But for Windows OS's there is can be not expected (likely - bad) results. VMware not much better in that plane (performance hit still there), Hyper-V native is the best choice for Win VM. But Hyper-V is not so good in advanced setup (can't passthrough USB, problems with PCIe passthrough for some devices and etc.).

 

[Anti-Ctrl]

I can agree, that Proxmox is very good (even - ideal) for UNIX-like systems virtualization. But for Windows OS's there is can be not expected (likely - bad) results. VMware not much better in that plane (performance hit still there), Hyper-V native is the best choice for Win VM. But Hyper-V is not so good in advanced setup (can't passthrough USB, problems with PCIe passthrough for some devices and etc.).

As far as i know (didnt test myself) ESXi has not a major Performance Penalty for Windows or Unix VMs. So its the best compromise from both worlds. I would really like to have someone from the Proxmox Staff to look into that, as its a major Problem in the Hypervisor. Cant be the case, that the same Hardware has much more Performance Potential on another HV. There must be something major wrong and it should be addressed.

 

[Gektor]

I think - it's a definitely bug of an benchmarking software. I have change processor type from host to kvm64 and get very weird results:

You need more tests in real work environment (not in benchmarks) to make sure that performance hit is not so big, as you think.

 

[Gektor]

I have change processor type to exactly architecture as on host (but not "host") - IvyBridge-IBRS:


IvyBridge:


Seems that benchmarking software has problems with results implementations or(and) some disagreement between systems with Spectre / Meltdown and other mitigations patches under Proxmox virtual environment...

P.S.
With "IvyBridge" in WinRAR i have get about 8600 KB/s (more close to bare metal and Hyper-V results).
With some tuning:

i can achieve ~8700 KB/s in WinRAR and AIDA64:

Selection between "host" and "correct architecture" - second variant is the best for benchmarks... But, as i write previously, in real work environment it can be different...

 

[MITM]

@Anti-Ctrl i was having the same benchmarking results as you in aida64 with X5675 (Westmere), report would be in the basement on memory speed,
but VM actually ran reasonable well (as far as i can tell never ran bare metal)
was following your posts here and on level1 looking for maximum performance.

so based on @Gektor findings I ran similar tests and changed the VM cpu settings - I was running host as recommended for performance.

• host (my base) = mem read 1,000mb/s (very low)
• westmere + nearly flags set= 15,212mb/s (15x faster)
• westmere+IRBS + nearly all flags set = 15,213mb/s
• westmere+IRBS + some flags set = 14,839mb/s

this also effected L1,2,3 cache similarly,
basically selecting correct architecture rather than host = much better aida64 benchmark.

attached some shots - upgraded from trial halfway through, re ran some test afterwards to confirm

Subjectively - this doesn't appear to effect usability and performance
Objectively - it does not effect CPU tests like cinebench (in VM I only get like 600 on r15 which I think its low)

if there is a simple test you are aware of that I can run in Linux vm, proxmox host and windows vm to compare performance
I can do that if it would help anyone.

 

[Anti-Ctrl]

@Anti-Ctrl i was having the same benchmarking results as you in aida64 with X5675 (Westmere), report would be in the basement on memory speed,
but VM actually ran reasonable well (as far as i can tell never ran bare metal)
was following your posts here and on level1 looking for maximum performance.

so based on @Gektor findings I ran similar tests and changed the VM cpu settings - I was running host as recommended for performance.

• host (my base) = mem read 1,000mb/s (very low)
• westmere + nearly flags set= 15,212mb/s (15x faster)
• westmere+IRBS + nearly all flags set = 15,213mb/s
• westmere+IRBS + some flags set = 14,839mb/s

this also effected L1,2,3 cache similarly,
basically selecting correct architecture rather than host = much better aida64 benchmark.

attached some shots - upgraded from trial halfway through, re ran some test afterwards to confirm

Subjectively - this doesn't appear to effect usability and performance
Objectively - it does not effect CPU tests like cinebench (in VM I only get like 600 on r15 which I think its low)

if there is a simple test you are aware of that I can run in Linux vm, proxmox host and windows vm to compare performance
I can do that if it would help anyone.


View attachment 25454

Thanks for testing that. Yall are right with that and i could replicate that. Still, the Performance in the Windows VM was sluggish and just overall slow. Cinebench was fine tho. I would like todo more testing, but tbh idk what todo anymore and i just slapped more money on the Problem and Bought new Hardware, as that problem persists now for over half a year i think. I´m atm in the Progress of restoring Backups of the VMs and reconfigureing my SMB Shares.

BTW: In CBR15 you should get around 800ish Points as far as i can tell. That being said, there where Spectre and Meltdown n stuff, so yeah, i guess 600 points is fine, as i only got around 1300 Point with TWO CPUs. So yeah, there is a bit lack of Power, but nothing i would consider mandatory.

For a Benchmark which runs on "everything", you should look into the Phoronix Test Suite. Its not that easy to setup on Windows, but most guides i found are fine and understandable. I think there are "only" three Memory tests which run on Windows, but that should be enough for comparison. I did that at one Point and had like 100x slower Performance in the Windows VM, than on Linux. And no I´m not being overdramatic here, it really was around 100x slower.

Thanks to all the Help that was provided here and especially thanks to you @MITM for offering more help on tests n Stuff. Unfortunately i already sold the Dual x5675 Board and, as I said, bought a new LGA 2011-3 Platform with a 2643 v3. My Initial Tests in the Windows VM (before reinstall) where good and it felt a thousand times better and faster. Pretty much native speed.

 

[_gabriel]

just hit the problem with AIDA64.
Switched to Windows integrated meomory benchmark , get correct values :
winsat mem

edit: Memory test from PerformanceTest of PassMark is correct too.