Re: [Qemu-devel] Re: [PATCH] Add cache=volatile parameter to -drive

[Alexander Graf]

Alexander Graf wrote:
> Anthony Liguori wrote:
>   
>> On 05/17/2010 08:17 AM, Alexander Graf wrote:
>>     
>>> On 17.05.2010, at 15:09, Anthony Liguori wrote:
>>>
>>>   
>>>       
>>>> On 05/17/2010 08:02 AM, Alexander Graf wrote:
>>>>     
>>>>         
>>>>>> My concern is that ext3 exaggerates the cost of fsync() which will
>>>>>> result in diminishing value over time for this feature as people
>>>>>> move to ext4/btrfs.
>>>>>>
>>>>>>          
>>>>>>             
>>>>> There will be ext3 file systems for years out. Just because people
>>>>> can use better and faster file systems doesn't mean they do. And
>>>>> I'm sure they can't always choose. If anything, I can try and see
>>>>> what the numbers look like for xfs.
>>>>>
>>>>>        
>>>>>           
>>>> But ext3 with barrier=1 is pretty uncommon in practice.  Another
>>>> data point would be an ext3 host file system with barrier=0.
>>>>      
>>>>         
>>> Who defines what is common and what not? To me, the SLES11 default is
>>> common. In fact, the numbers in the referred mail were done on an
>>> 11.1 system.
>>>    
>>>       
>> But it wasn't the SLES10 default so there's a smaller window of
>> systems that are going to be configured this way.  But this is
>> orthogonal to the main point.  Let's quantify how important this
>> detail is before we discuss the affected user base.
>>     
>
> Alright. I took my Netbook (2GB of RAM) and a USB hard disk, so I can
> easily remount the data fs the vmdk image is on. Here are the results:
>
> # mkfs.ext3 /dev/sdc1
> # mount /dev/sdc1 /mnt -obarrier=1
>
> cache=writeback
>
> real    0m52.801s
> user    0m16.065s
> sys     0m6.688s
>
> cache=volatile
>
> real    0m47.876s
> user    0m15.921s
> sys     0m6.548s
>
> # mount /dev/sdc1 /mnt -obarrier=0
>
> cache=writeback
>
> real    0m53.588s
> user    0m15.901s
> sys     0m6.576s
>
> cache=volatile
>
> real    0m48.715s
> user    0m16.581s
> sys     0m5.856s
>
> I don't see a difference between the results. Apparently the barrier
> option doesn't change a thing.
>   

The same test case for XFS:

cache=writeback

real    0m50.868s
user    0m11.133s
sys    0m12.733s

cache=volatile

real    0m43.680s
user    0m16.089s
sys    0m7.812s

Though I did have numbers here going as far down as 25 seconds for a run!


Alex