Memory in vm half as fast after migration of vm.

Darkbeldin

@andreas Ok so after discussing it with Dev team the issue has been identified.
The trouble is linked to TSC management in the VM.
You can work around the issue by setting the VM:

xe vm-param-set uuid=<VM_UUID> platform:tsc_mode=2

But be aware we can not recommend this settings to go to a production VM.
TSC clock won't be emulated at all if you enable this settings. So you might have some weird time behavior during migration.

Andreas

@darkbeldin Okay thanks
I did test this and it worked.

TheNorthernLight

@darkbeldin Hello, So for those of us in production, does this problem affect rolling pool upgrades?

If so, how do we fix this and update our pools without needing to explicitly shutdown VMs in the pool?

Andrew

@darkbeldin Is memory access actually slower or is a timing issue with the statistics?

Darkbeldin

@andrew Sorry guys not sure i understand the issue enough to answer, @olivierlambert has a way better understanding of it, think it's better you to ask him

olivierlambert

It's a very long story. The real impact isn't that big in real usage, and it depends on so many factors that it's hard to really know at one time if you are really affected or not.

The core issue is related to TSC clock. Time/tick regularity on hardware is a REAL mess, even on the same hardware, Xen default mode is trying to use the TSC without emulation for your VMs, but sometimes TSC is doing weird things, and Xen is able to preserve the behavior in the guest by emulating it.

This emulation is costing performance. And this is already on the very same hardware. Now imagine live migrate to another machine, to another CPU and motherboard, even on the exact same model. The TSC frequency can't be exactly the same, so there's some variation.

To keep a perfectly constant/consistent clock on the VM, Xen default TSC mode (1) is detecting those changes to "hide" them to the guest with some emulation (if needed).

Mode 2 is "no emulation whatsoever" (and mode 0 is always emulate). I'm not exactly sure about the risk on switching to mode 2 in production. If you want to test it and check chrony/ntpd logs, I'm interested in the results

olivierlambert

Here is an old paper from VMware, but with a good recap on various timers and the complexity of it: https://nextcloud.vates.fr/index.php/s/WHk64gHTK4iaJAP

Forza

@olivierlambert said in Memory in vm half as fast after migration of vm.:

Mode 2 is "no emulation whatsoever" (and mode 0 is always emulate). I'm not exactly sure about the risk on switching to mode 2 in production. If you want to test it and check chrony/ntpd logs, I'm interested in the results

We use local NTP servers so we could use chrony the the like to sync (already do). But, is TSC only about timesync, or is it about other stuff like Linux kernel internals depending on some stability of it?
I found this article https://superuser.com/questions/393969/what-does-clocksource-tsc-unstable-mean that discusses TSC a little. It seems we can have unstable TSC on multicore systems and the kernels should handle it anyway?

olivierlambert

Frankly, I wouldn't speculate on the potential risk, I prefer to answer that I don't know I might read stuff later when I can to get a better idea.

Forza

@olivierlambert said in Memory in vm half as fast after migration of vm.:

Frankly, I wouldn't speculate on the potential risk, I prefer to answer that I don't know I might read stuff later when I can to get a better idea.

And.. Why does the emulation make it slower after a move - is this a bug in the emulation?

olivierlambert

No it's not a bug. Emulating a TSC clock is taking resources. Emulation in general is bad for performance.

Forza

@olivierlambert said in Memory in vm half as fast after migration of vm.:

No it's not a bug. Emulating a TSC clock is taking resources. Emulation in general is bad for performance.

So what happens is that no emulation is used when the VM is booted, but it gets activated when migrating to another host?

And not enabling TSC emulation would mean that TSC behaviour/properties might change when migrating, which in turn has a possible (what?) effect on the guest VM?

olivierlambert

That's correct. Because of TSC variation between different platforms
Correct. It means the guest will see that, and the unknown to me is the real effect/risk on the guest

Forza

@olivierlambert said in Memory in vm half as fast after migration of vm.:

That's correct. Because of TSC variation between different platforms

Correct. It means the guest will see that, and the unknown to me is the real effect/risk on the guest

Thanks for helping clarifying the problem. Definitely something worth testing fully before making changes.

olivierlambert

We won't change the default behavior before having a LARGE use base using it without any problem And if I wonder about the risk on Linux, I have 0 knowledge on Windows based OS.