Epyc VM to VM networking slow
-
This is already different result than @nicols
-
@olivierlambert Not really...
In Nicols first post, a single threaded iperf got 3.38Gb/s.
Single threaded for me was 3.32Gb/s.With two threads I get 5.19Gb/s
With 20 threads I cap off at 7.02Gb/sThis performance is about the same with Windows VM's and Debian VM's. So it's not a guest OS issue. It's something in the hypervisor.
-
A note...
I'm running a single 16 core, 32 thread second-gen Epyc.
Nicols is running a dual proc, 24 core, 48 thread third-gen Epyc.My base clock rate is 3.0Ghz. His is 2.9Ghz.
The improved caching and memory handling in the third-gen Epyc should be behaving better than my second gen CPU's, but generally speaking, our performance seems to be the same.
-
He said he could reach 18gbits between 2x Windows VM, if I remember correctly.
I wonder about the guest kernel too (Debian 11 vs 12)
-
@olivierlambert With a billion threads.
Anyway...
I'm most definitely a willing subject to help get this resolved. Heck..I'll even give you guys access to the environment to do whatever you want to do. I would just like to see this get fixed.
With that...You guys tell me. What tests do you want run and do you want access to the environment to do your own thing with it?
-
Can you reproduce the same speed he got on your side? Because first we need a consistent result between multiple people to be sure it's not platform related (ie Supermicro or something).
The only reason we could think it's not normal is the difference between Intel and AMD, that shouldn't be that huge. Or maybe AMD CPU are a lot slower with
memcpy()?
-
@JamesG said in Epyc VM to VM networking slow:
@olivierlambert With a billion threads.
nope, Win 10 VM with 4 or 8 VCPU and 8GB RAM
But, with billion threads in Linux VM, speed increases up to 8 threats, then it drops again. -
@olivierlambert For single threaded iperf....Yes. Our speeds match 100%. Which is half the transfer rate of a single threaded iperf on 12 year-old Xeon E3 hardware.
I understand that we've had lots of security issues in the past decade and several steps have been taken to protect and isolate the memory inside all virtualization platforms. When I first built my E3-1230 Xeon system for homelab, VM to VM iperfs were like 20Gb/s. Nowadays that's significantly slowed down.
Anyway...I just find it hard to believe that with as superior a computing platform as Epyc is, that the single-threaded iperf is so much slwoer than 12-year-old entry level Intel CPUs.
Maybe I should load VMWare on this system and see how it does and report back. Same hardware, but different hypervisor, and compare notes.
-
@nicols said in Epyc VM to VM networking slow:
@JamesG said in Epyc VM to VM networking slow:
@olivierlambert With a billion threads.
nope, Win 10 VM with 4 or 8 VCPU and 8GB RAM
But, with billion threads in Linux VM, speed increases up to 8 threats, then it drops again.This is with 1 and 16 threads:
-
@nicols said in Epyc VM to VM networking slow:
@nicols said in Epyc VM to VM networking slow:
@JamesG said in Epyc VM to VM networking slow:
@olivierlambert With a billion threads.
nope, Win 10 VM with 4 or 8 VCPU and 8GB RAM
But, with billion threads in Linux VM, speed increases up to 8 threats, then it drops again.This is with 1 and 16 threads:
also, this:
https://nextcloud.openit.hr/s/CptZpTt4jbWcRPX
is cpu load on host during 2 VM linux doing 16 thread iperf (with cumulative speed of pathetic 4 Gbit/sec).
It seems way to high for this kind of job? -
@olivierlambert said in Epyc VM to VM networking slow:
Or maybe AMD CPU are a lot slower with memcpy()?
Has anyone reviewed this issue? Is there a way to test with a newer version of
glibc? -
@nicols give me your VM specs and I'll run the exact same tests. vCPU, RAM, anything else relevant.
-
@Danp That's interesting...
-
@JamesG said in Epyc VM to VM networking slow:
@nicols give me your VM specs and I'll run the exact same tests. vCPU, RAM, anything else relevant.
Debian 12: 16 VCPU, 2GB RAM
Windows 10 pro: 16 VCPU, 8GB RAM, citrix vmtols 9.3.1On Linux Debian there is no much difference between 8 and 16 VCPU
On Windows 10, 8 VCPU: 16 Gbit/sec, 16 VCPU: 21 Gbit/sec -
@JamesG said in Epyc VM to VM networking slow:
@Danp That's interesting...
Yes, it is, but as i wrote earlier, i get full 21 Gbps Linux VM to VM on Proxmox/KVM (on exact same host, same BIOS settings), so i think it must be some problem on relation Epyc - Xen hypervisor....
-
@nicols Agreed. I'm pretty sure this is a Xen/Epyc issue.
This evening I'll build a couple of VM's to your config, run iperf, and report back the results.
-
@nicols said in Epyc VM to VM networking slow:
i get full 21 Gbps Linux VM to VM on Proxmox/KVM
If
glibcis the source of the issue, then a likely explanation for your results is that Proxmox/KVM are using an updated version of this library where the patch has been applied.@olivierlambert Do you know if anyone on your team has looked into this?
-
We are very very busy ATM.
Also, about comparing to KVM doesn't make sense at all: there's no such network/disk isolation in KVM, so you can do zero copy, which will yield to much better performances (at the price of the thin isolation).
First, we should compare between 2x fully patched systems (one Intel one AMD) a similar config, we could have a baseline and understands why AMD is a lot slower.
-
Adding @dthenot in the loop in case it rings a bell.
-
The past couple of days have been pretty nuts, but I've dabbled with testing this and in my configuration with XCP-ng 8.3 with all currently released patches, I top out at 15Gb/s with 8 threads on Win 10. Going further to 16 threads or beyond doesn't really improve things.
Killing core boost, SMT, and setting deterministic performance in BIOS added nearly 2Gb/s on single-threaded iperf.
When running the iperf and watching htop on the XCP-ng server, I see nearly all cores running at 15-20% for the duration of the transfer. That seems excessive.
Iperf on the E3-1230v2...Single thread, 9.27Gbs. Neglibile improvement for more threads. Surprisingly, a similar hit on CPU performance. Not as bad though. 10Gbps traffic hits about 10% or so. Definitely not as bad as on the Epyc system.
I'll do more thorough testing tomorrow.
