Keeping a guest cpu busy actually makes the guest faster?

[940607]

Hi.
I have Arch64 guest on Win7 host. Main problem is: it mostly takes 2-5 seconds to rewrite a 100kb file residing on the guest's samba share. I also tried to upload and download the files between host and guest and the speed is unstable.

Surprisingly, running this shell command on guest in background workarounds the problem:

Code: Select all

for ((i=0;;i++)); do : ; done

This is my work PC. I also have a laptop with a similar setup and there's no such problem with it. Can it be some power saving misbehavior? Where to dig?

Code: Select all

- The version of VirtualBox you are using, and whether you have installed the Guest Additions (in the guest).
5.2.8 r121009 / no additions
- Host & Guest make and version including 32 or 64 bit, and the amount of memory available to both.
Host: Win7 Pro 64 SP1
  RAM: 24GB
Guest: Linux archlinux 4.15.9-1-ARCH #1 SMP PREEMPT Sun Mar 11 17:54:33 UTC 2018 x86_64 GNU/Linux
  RAM: 4 GB

[socratis]

I'd start digging into the differences between your two hosts and/or the recipes for the guests. I don't have either of them in front of me, so I can't really tell you what the differences might be.

[michaln]

Yes, it could be power saving (doesn't have to be, but could be). It's not "misbehavior", it's "working as designed". It is typical that when a CPU is not fully utilized, it runs slower. Keeping it busy forces it to run at full speed.

[940607]

Looks like I suffered from the odd number of vCPUs (3 in my case). I followed the FAQ advice to keep one core to the host and this is what happened.

[940607]

Nope. Two vCPUs didn't solve the problem.

[socratis]

I followed the FAQ advice

What FAQ advice? We haven't even seen a VBox.log from you...

[940607]

The FAQ here : viewtopic.php?f=35&t=77413
says:

always keep a core ( not threads ) for the host so the host side code and host itself does not become over burdened

But again, reducing the cores from 3 to 2 didn't help. There are signs that with just one core network speed is stable.

I attached the logs to the opening post. During the test I was sending zeros to the xinetd discard service on guest on port 9. The guest network adapter is a host-only adapter.

Code: Select all

>/dev/tcp/arch64/9 pv /dev/zero

The pv tool was showing the current upload speed. While I was doing nothing else the speed was around 2 MiB/s. However, launching another upload made the sum of speeds grow to 140 MiB/s.

[michaln]

This is probably something related to scheduling. It is a well known problem that some workloads run slower with 2 CPUs than with one (and the effect is much worse in VMs). It happens with client/server workloads where the client generates a large number of requests for the server. With multiple CPUs, this ends up bouncing the work across CPUs and that just slows things down.

What probably happens is that when you keep one CPU busy in the guest, the workload you're interested in is forced to run on the other CPU, and it runs faster.