VM guest tuning?

[Hector Ordorica]

First of all, thanks for BFQ, I can really notice the improvement in system latency. My system doesn't lock up during heavy compiles or disk i/o usage.

It seems to work great for all applications in my Ubuntu linux with kernel 3.13.0-rc7+. However, I'm also running VMWare workstation 10 with a Windows 7 guest that takes much longer to boot from a cold cache, and the applications (like Outlook 2010) take a significantly longer time to open compared to deadline.

Now perhaps BFQ is doing it's job, because the rest of my Linux system remains responsive while Windows is working in the background. But should Windows take that much of a performance hit?

I was wondering if it was because all BFQ can see is VMWare reading and writing I/O and can't peer into what Windows is doing.

I'm running a 128GB samsung PM830 SSD, with XFS.

Thanks. Let me know if there's anything I can tune for this.

[Paolo Valente]

Il giorno 14/gen/2014, alle ore 23:04, Hector Ordorica ha scritto:

> First of all, thanks for BFQ,

You're welcome (also as a BFQ user :) ).

> I can really notice the improvement in system latency. My system doesn't lock up during heavy compiles or disk i/o usage.

Definitely glad to hear that.

>
> It seems to work great for all applications in my Ubuntu linux with kernel 3.13.0-rc7+. However, I'm also running VMWare workstation 10 with a Windows 7 guest that takes much longer to boot from a cold cache, and the applications (like Outlook 2010) take a significantly longer time to open compared to deadline.
>
> Now perhaps BFQ is doing it's job, because the rest of my Linux system remains responsive while Windows is working in the background. But should Windows take that much of a performance hit?
>
> I was wondering if it was because all BFQ can see is VMWare reading and writing I/O and can't peer into what Windows is doing.

You described both the symptoms and the causes perfectly.

About one month ago I submitted a paper describing this very problem and its root cause: the fact that the host I/O scheduler may have no hint on what is happening inside the guest OS, and, in particular, on when the guest I/O needs to be privileged to reduce latency in the guest itself. I am also working on a solution to this problem, as the main goal of a contract with a french enterprise named VoSys.
The solution we devised so far requires a form of coordination between the I/O schedulers in guest and in the host. Hence, in practice, BFQ needs to be extended. And we are extending it for QEMU/KVM.
I think this solution might work, maybe after some minor modifications, also for other VM emulators.

This means however that, if the guest OS is Windows, and hence no BFQ scheduler is available, this contribution will be useless for your case …

As for deadline, probably it treats the VMware I/O better than BFQ exactly for the reason you said: BFQ probably considers the I/O pattern generated by VMWare as non interactive, and hence reduces the throughput reserved for this I/O, to privilege other processes (correctly) deemed as interactive.

>
> I'm running a 128GB samsung PM830 SSD, with XFS.
>
> Thanks. Let me know if there's anything I can tune for this.
>

As a possible tuning attempt, you might try to assign, manually, a higher I/O weight to the VMware processes in charge of doing I/O for the guest OS. To this purpose, you may use either ionice or the cgroups interface. As possible weight values, try 10 or even 20. I am not sure about the effectiveness and the side effects of such a manual intervention. If you try, let us know.

Finally, perhaps the next v7 version of BFQ might mitigate the important problem you reported. But also in this case I am not sure.

Paolo


>
>
> --
> You received this message because you are subscribed to the Google Groups "bfq-iosched" group.
> To unsubscribe from this group and stop receiving emails from it, send an email to bfq-iosched...@googlegroups.com.
> For more options, visit https://groups.google.com/groups/opt_out.