Linux Virtualization - Nov 2018 - [PATCH 0/1] vhost: add vhost_blk driver

On 2018/11/5 ??11:23, Vitaly Mayatskih wrote:
> On Sun, Nov 4, 2018 at 10:00 PM Jason Wang <jasowang at redhat.com>
wrote:
>
>>> # fio num-jobs
>>> # A: bare metal over block
>>> # B: bare metal over file
>>> # C: virtio-blk over block
>>> # D: virtio-blk over file
>>> # E: vhost-blk bio over block
>>> # F: vhost-blk kiocb over block
>>> # G: vhost-blk kiocb over file
>>> #
>>> #  A     B     C    D    E     F     G
>>> 16 1480k 1506k 101k 102k 1346k 1202k 566k
>> Hi:
>>
>> Thanks for the patches.
>>
>> This is not the first attempt for having vhost-blk:
>>
>> - Badari's version: https://lwn.net/Articles/379864/
>>
>> - Asias' version: https://lwn.net/Articles/519880/
>>
>> It's better to describe the differences (kiocb vs bio?
performance?).
>> E.g if my memory is correct, Asias said it doesn't give much
improvement
>> compared with userspace qemu.
>>
>> And what's more important, I believe we tend to use virtio-scsi
nowdays.
>> So what's the advantages of vhost-blk over vhost-scsi?
> Hi,
>
> Yes, I saw both. Frankly, my implementation is not that different,
> because the whole thing has only twice more LOC that vhost/test.c.
>
> I posted my numbers (see in quoted text above the 16 queues case),
> IOPS goes from ~100k to 1.2M and almost reaches the physical
> limitation of the backend.
>
> submit_bio() is a bit faster, but can't be used for disk images placed
> on a file system. I have that submit_bio implementation too.
>
> Storage industry is shifting away from SCSI, which has a scaling
> problem.

Know little about storage. For scaling, do you mean SCSI protocol 
itself? If not, it's probably not a real issue for virtio-scsi itself.

>   I can compare vhost-scsi vs vhost-blk if you are curious.

It would be very helpful to see the performance comparison.


Thanks

> Thanks!

On Tue, Nov 06, 2018 at 10:45:08AM +0800, Jason Wang
wrote:
> > Storage industry is shifting away from SCSI, which has a scaling
> > problem.
> 
> 
> Know little about storage. For scaling, do you mean SCSI protocol itself?
If
> not, it's probably not a real issue for virtio-scsi itself.
The above is utter bullshit.  There is a big NVMe hype, but it is not
because "SCSI has a scaling problem", but because the industry can
sell
a new thing, and the standardization body seems easier to work with.

On 06/11/2018 08:13, Christoph Hellwig wrote:
> On Tue, Nov 06, 2018 at 10:45:08AM +0800, Jason Wang wrote:
>>> Storage industry is shifting away from SCSI, which has a scaling
>>> problem.
>>
>> Know little about storage. For scaling, do you mean SCSI protocol
itself? If
>> not, it's probably not a real issue for virtio-scsi itself.
> 
> The above is utter bullshit.  There is a big NVMe hype, but it is not
> because "SCSI has a scaling problem", but because the industry
can sell
> a new thing, and the standardization body seems easier to work with.
In the case of Linux, there is indeed some performance to be gained by
skipping the SCSI layer in the guest and to a lesser extent in QEMU,
which is why virtio-blk is still in use.  But there's no scaling
problem, it's just extra constant overhead.

Paolo