Using tools such as ionice inside guests

Using tools such as ionice inside guests

[Emmanuel Lacour]

Dear kvm developpers/users,

I just saw that a change has been made in the kernel so the only one
scheduler available in guests for virtio-blk is now "none".

Seems introduced by [1].

I tried to used ionice on a 3.16 guest with virtio-blk and in fact it no
longer works because cfq is no longer available.

Is there a way to use, inside a guest, a tool like ionice to set
different IO priority in guests processes? It's very useful for loering
down the performance impact of backups jobs for example.




[1] -
http://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/commit/?id=a98755c559e0e944a44174883b74a97019e3a367

Re: Using tools such as ionice inside guests

[Paolo Bonzini]

On 18/05/2016 17:49, Emmanuel Lacour wrote:
> Dear kvm developpers/users,
> 
> I just saw that a change has been made in the kernel so the only one
> scheduler available in guests for virtio-blk is now "none".
> 
> Seems introduced by [1].

Not by this one; rather this one:

http://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/commit/?id=1cf7e9c68fe84248174e998922b39e508375e7c1

> I tried to used ionice on a 3.16 guest with virtio-blk and in fact it no
> longer works because cfq is no longer available.
> 
> Is there a way to use, inside a guest, a tool like ionice to set
> different IO priority in guests processes? It's very useful for loering
> down the performance impact of backups jobs for example.
> 
> [1] -
> http://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/commit/?id=a98755c559e0e944a44174883b74a97019e3a367
> --

Not anymore unfortunately.  The effect of this change was not
advertised enough, in my opinion.  The blk-mq path should have replaced
the code used by use_bio=1, but not the code used by use_bio=0.

Thanks,

Paolo

Re: Using tools such as ionice inside guests

[Emmanuel Lacour]

Thanks, it's a bad news for us :(

I will try to play with cgroups io throttling to limit by iops or bytes
read, but it'll be a hard limit rather than a priority :(

Do you think there could be back some io priority throttling in not so
far future kernels?

Re: Using tools such as ionice inside guests

[Paolo Bonzini]

On 18/05/2016 18:26, Emmanuel Lacour wrote:
> Thanks, it's a bad news for us :(
> 
> I will try to play with cgroups io throttling to limit by iops or bytes
> read, but it'll be a hard limit rather than a priority :(

That probably won't work either, but you can use io throttling on the
host too.

> Do you think there could be back some io priority throttling in not so
> far future kernels?

I wouldn't oppose a patch to add back the non-mq path, but as far as I
know nobody is working on it.  It's not hard and I can help if you need
guidance.

Adding I/O scheduler support to blk-mq has been promised for at least a
year now, but I'm not aware of which kernel release might have the work.
 It's not even been submitted to LKML, so I guess it's quite far away.

Thanks,

Paolo

Re: Using tools such as ionice inside guests

[Christian Borntraeger]

On 05/18/2016 07:12 PM, Paolo Bonzini wrote:
> 
> 
> On 18/05/2016 18:26, Emmanuel Lacour wrote:
>> Thanks, it's a bad news for us :(
>>
>> I will try to play with cgroups io throttling to limit by iops or bytes
>> read, but it'll be a hard limit rather than a priority :(
> 
> That probably won't work either, but you can use io throttling on the
> host too.
> 
>> Do you think there could be back some io priority throttling in not so
>> far future kernels?
> 
> I wouldn't oppose a patch to add back the non-mq path, but as far as I
> know nobody is working on it.  It's not hard and I can help if you need
> guidance.
> 
> Adding I/O scheduler support to blk-mq has been promised for at least a
> year now, but I'm not aware of which kernel release might have the work.
>  It's not even been submitted to LKML, so I guess it's quite far away.

What actually happened to the virtio-blk multiqueue support in QEMU?

Re: Using tools such as ionice inside guests

[Paolo Bonzini]

On 19/05/2016 22:12, Christian Borntraeger wrote:
> On 05/18/2016 07:12 PM, Paolo Bonzini wrote:
>>
>>
>> On 18/05/2016 18:26, Emmanuel Lacour wrote:
>>> Thanks, it's a bad news for us :(
>>>
>>> I will try to play with cgroups io throttling to limit by iops or bytes
>>> read, but it'll be a hard limit rather than a priority :(
>>
>> That probably won't work either, but you can use io throttling on the
>> host too.
>>
>>> Do you think there could be back some io priority throttling in not so
>>> far future kernels?
>>
>> I wouldn't oppose a patch to add back the non-mq path, but as far as I
>> know nobody is working on it.  It's not hard and I can help if you need
>> guidance.
>>
>> Adding I/O scheduler support to blk-mq has been promised for at least a
>> year now, but I'm not aware of which kernel release might have the work.
>>  It's not even been submitted to LKML, so I guess it's quite far away.
> 
> What actually happened to the virtio-blk multiqueue support in QEMU?

The guy who submitted the patch for the "fake" virtio-blk multiqueue
(backed by a single iothread) disappeared.  "Real" virtio-blk multiqueue
is at least 80 patches away (70 of those have been written :)).

However, multiqueue is independent from Emmanuel's problem.

Thanks,

Paolo

Re: Using tools such as ionice inside guests

[Emmanuel Lacour]

Le 18/05/2016 19:12, Paolo Bonzini a écrit :
> 
> 
> On 18/05/2016 18:26, Emmanuel Lacour wrote:
>> Thanks, it's a bad news for us :(
>>
>> I will try to play with cgroups io throttling to limit by iops or bytes
>> read, but it'll be a hard limit rather than a priority :(
> 
> That probably won't work either, but you can use io throttling on the
> host too.
> 

I'll try and let you now if it works. I cannot do this on the host,
because it's per processes inside each VMs, not betweeen VMs.

>> Do you think there could be back some io priority throttling in not so
>> far future kernels?
> 
> I wouldn't oppose a patch to add back the non-mq path, but as far as I
> know nobody is working on it.  It's not hard and I can help if you need
> guidance.
> 
> Adding I/O scheduler support to blk-mq has been promised for at least a
> year now, but I'm not aware of which kernel release might have the work.
>  It's not even been submitted to LKML, so I guess it's quite far away.
> 

Unfortunatly, I'm a sysadmin with too much little dev knowledge :(