RE: [Patch v4 0/3] Introduce a driver to support host accelerated access to Microsoft Azure Blob

[Long Li]

> Subject: Re: [Patch v4 0/3] Introduce a driver to support host accelerated
> access to Microsoft Azure Blob
> 
> On Mon, Jul 19, 2021 at 08:31:03PM -0700, longli@xxxxxxxxxxxxxxxxx wrote:
> > From: Long Li <longli@xxxxxxxxxxxxx>
> >
> > Microsoft Azure Blob storage service exposes a REST API to
> > applications for data access.
> >
> (https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fdoc
> > s.microsoft.com%2Fen-us%2Frest%2Fapi%2Fstorageservices%2Fblob-
> service-
> > rest-
> api&amp;data=04%7C01%7Clongli%40microsoft.com%7Ce499fbe161554232e
> >
> b1608d94b96a772%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C637
> 623932
> >
> 843247787%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIj
> oiV2luMzIi
> >
> LCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C2000&amp;sdata=9CKNHAmurdBWp
> ZeLfkiC18
> > CXNg66UhKsSZzzHZkzf0Y%3D&amp;reserved=0)
> >
> > This patchset implements a VSC (Virtualization Service Consumer) that
> > communicates with a VSP (Virtualization Service Provider) on the
> > Hyper-V host to execute Blob storage access via native network stack on
> the host.
> > This VSC doesn't implement the semantics of REST API. Those are
> > implemented in user-space. The VSC provides a fast data path to VSP.
> >
> > Answers to some previous questions discussing the driver:
> >
> > Q: Why this driver doesn't use the block layer
> >
> > A: The Azure Blob is based on a model of object oriented storage. The
> > storage object is not modeled in block sectors. While it's possible to
> > present the storage object as a block device (assuming it makes sense
> > to fake all the block device attributes), we lose the ability to
> > express functionality that are defined in the REST API.
> 
> What "REST API"?
> 
> Why doesn't object oriented storage map to a file handle to read from?
> No need to mess with "blocks", why would you care about them?
> 
> And again, you loose all caching, this thing has got to be really slow, why add
> a slow storage interface?  What workload requires this type of slow block
> storage?

"Blob REST API" expresses storage request semantics through HTTP. In Blob
REST API, each request is associated with a context meta data expressed in
HTTP headers. The ability to express those semantics is rich, it's only limited
by HTTP protocol.

There are attempts to implement the Blob as a file system.
Here is an example filesystem (BlobFuse) implemented for Blob:
(https://github.com/Azure/azure-storage-fuse).

It's doable, but at the same time you lose all the performance and
shareable/security features presented in the REST API for Blob. A POSIX
interface cannot express same functionality as the REST API for Blob.

For example, The Blob API for read (Get Blob, 
https://docs.microsoft.com/en-us/rest/api/storageservices/get-blob)
has rich meta data context that cannot easily be mapped to POSIX. The same
goes to other Blob API to manage security tokens and the life cycle of shareable
objects.

BlobFuse (above) filesystem demonstrated why Blob should not be implemented
on a filesystem. It's useable for data management purposes. It's not usable for an I/O
intensive workload. It's not usable for managing sharable objects and security tokens.

Blob is designed not to use file system caching and block layer I/O scheduling.
There are many solutions existing today, that access raw disk for I/O, bypassing
filesystem and block layer. For example, many database applications access raw
disks for I/O. When the application knows the I/O pattern and its intended behavior,
it doesn't get much benefit from filesystem or block.

> 
> > Q: You also just abandoned the POSIX model and forced people to use a
> > random-custom-library just to access their storage devices, breaking
> > all existing programs in the world?
> >
> > A: The existing Blob applications access storage via HTTP (REST API).
> > They don't use POSIX interface. The interface for Azure Blob is not
> > designed on POSIX.
> 
> I do not see a HTTP interface here, what does that have to do with the kernel?
> 
> I see a single ioctl interface, that's all.

The driver doesn't attempt to implement Blob API or HTTP. It presents a fast data
path to pass Blob requests to hosts, so the guest VM doesn't need to assemble
a HTTP packet for each Blob REST requests. This also eliminates additional
overhead in guest network stack to send the HTTP packets over TCP/IP.

Once the request reaches the Azure host, it knows the best way to reach to the 
backend storage and serving the Blob request, while at the same time all the 
security and integrity features are preserved.


> 
> > Q: What programs today use this new API?
> >
> > A: Currently none is released. But per above, there are also none
> > using POSIX.
> 
> Great, so no one uses this, so who is asking for and requiring this thing?
> 
> What about just doing it all from userspace using FUSE?  Why does this HAVE
> to be a kernel driver?

We have a major workload nearing the end of development. Compared with
REST over HTTP, this device model presented faster data access and CPU savings
in that there is no overhead of sending HTTP over network.

Eventually, all the existing Blob REST API applications can use this new API, once
it gets to their Blob transport libraries.

I have explained why BlobFuse is not suitable for production workloads. There
are people using BlobFuse but mostly for management purposes.

Thanks,
Long

> 
> thanks,
> 
> greg k-h