Re: [Qemu-discuss] A question about virtio inside qemu

[Stefan Hajnoczi]

On Wed, Dec 11, 2013 at 08:55:33PM +0000, Yaodong Yang wrote:
> I have a quick question about the virtio inside qemu. When the user 
> application sends to a specific virtual disk a large number of read requests 
> in a very short time, where should these requests be queued? Inside the 
> virtqueue? virtqueue available ring or the underlying bdrv_read_aio()?

Requests take up descriptor space in the vring.  Remember the vring
consists of 3 things:
1. Descriptor space
2. Available ring
3. Used ring

While the request is pending its vring descriptors will be in use,
regardless of whether the head index is in the avail/used rings.

> Currently, I run one guest vm with two virtual disks (one for ubuntu system, 
> the other is to serve the data access.) When the system is running, I can 
> check there are for instance 50 inflight requests to read or write the system 
> virtual disk. However, the inflight requests to the data virtual disk is 
> always one. Which is not reasonable, because inside the vm we calculated the 
> avery response time for each request is much larger than the time interval 
> between two adjacent requests.

I don't understand this paragraph.  But virtio-blk and virtio-scsi
support multiple in-flight requests.  The guest should be able to
achieve queue depths of around 128 with virtio-blk.

> My ultimate goal is to get the number of in-flight requests for a certain 
> virtual disk. Could someone give me some hints?

Tools like iostat(1) can tell you about in-flight requests at the Linux
block layer level.

Stefan