Cloud Native Computing Foundation KubeCon + CloudNativeCon China 2019 (Shanghai), 5 Jul 2019

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Introduce an SPDK vhost FUSE Target to Accelerate File Access in... Liu Changpeng & Liu Xiaodongl

Description

Join us for Kubernetes Forums Seoul, Sydney, Bengaluru and Delhi - learn more at kubecon.io

Don't miss KubeCon + CloudNativeCon 2020 events in Amsterdam March 30 - April 2, Shanghai July 28-30 and Boston November 17-20! Learn more at kubecon.io. The conference features presentations from developers and end users of Kubernetes, Prometheus, Envoy, and all of the other CNCF-hosted projects

Introduce an SPDK vhost FUSE Target to Accelerate File Access in VM and Containers - Liu Changpeng & Liu Xiaodongl, Intel

Virtio-fs(https://virtio-fs.gitlab.io/) is proposed recently to provide file system sharing for lightweight VMs and containers workloads, where shared volumes are a requirement. In this presentation, we propose an SPDK(Storage Performance Development Kit, https://spdk.io) userspace vhost-user-fs solution, which can be used together with QEMU/Kata Container to accelerate virtio-fs. Virtio-fs uses FUSE instead of 9P for communication. We will present this solution in details including the utilization of techniques such as virtio-fs, blobfs (SPDK file system) and the significant performance gain achieved. Blobfs can be built on abstract block device layer in SPDK, which can access local or remote storage services via iSCSI/NVMe/NVMeoF protocols in userspace. Relying on this solution, we are going to build a fast, consistent and secure manner to share directory tree on host to guests.

https://sched.co/Nrv9

A

B

Everyone I'm engineer from Intel. My name is Duchamp owns. This is my colleague goes out on, so we have achieved SPD kV host fuse target I'm going to present the slides.

A

B

We do these things, we want to tell you, why do you want to create and we host fuse target if you listen to the morning session or if you have used to container or VM for 9 p FS, maybe you will be more familiar actually in the beginning, if want to use the user stage where you using 9p, we were also discussing about the disadvantages of 9p.

B

We have also talked about whatever FS. So then we ship to fuse. We have already do something, and we have also done a simple compatibility which feels target. That's. Why I want to demonstrate that to you about what we have achieved.

A

B

We start with this topic: I want to tell you what is as PDK actually is instead of libraries at the bottom is the qemu, and then these are the blocks and storage layers and I want to highlight about what we're doing is up to 2 million as a call, so we really reached tens of millions cause levels on our webpage. We have described about what other things we have done and we'll also have in comparison with the new technology.

A

B

Is the architecture of SP DK from the one layer we have enemy and also we have quick data and copy engine and based on that, we have storage services and we have block device dry.

B

Larger volume can also be used to flow block management. We also have the crypto. We also have calcium cache, and this is the open channel FTL. This is compression.

A

B

That we have been storage protocols, nvme target.

A

B

The previous topic we have already talked about that in the V host. We have we host enemy. We host today we're talking about the host first target, and this is a user stage.

B

Implementation.

A

B

We also have block store, which is a block management distributor.

A

B

One is about unified, SP dkv host user target.

B

We can look at this. We support the host user block, we support what house Gazi and similarly, we support the host.

B

So we can use V, mu, G.

A

B

To decipher- and this is about the the host issues target.

B

So we can build on comedy host user library if we don't know about the host.

B

Actually, what how is define is back. We host this a time limit. Basically, there are two experiments, colonel and.

A

In colonel and.

B

Also, the host use.

A

B

The V host target the biggest advantage of the host is that eliminates a man, my all access for submission with Pauling mechanism.

B

And the way to achieve that is to use the polling mechanism.

A

B

We're sharing a file from gas to host our existing.

B

Protocol is 9 PFS. There are also two ways to achieve that. I have listed here.

A

B

First, one is this: what album 9:00 p.m.

B

If you use keeled process, our processing would be within the Kim. You.

A

B

This is what how 9p- and there is also the house 9p. I didn't add, user' here, because this is a kernel solution.

A

B

The kernels Lucia there is also upstream minimum strong.

A

B

This is the V host 9p kernel.

A

B

Let's look at the backend for the front and it's a saying we have 9p Colonel. We have 9 P driver for the backend.

B

We use what house 9 p to send to request, and then we use the standard port to cold out to the standard system from front to back end.

A

B

Talking about what's a fuse, it's actually file system in userspace, they consist of two parts. One is.

A

B

Driver we at this virtual file system is a common layer and Linux, and the first driver is a kernel space and the other one is Libre fuse.

A

B

A

B

Lip use and they will be able to achieve a customized fewest demo if we're going to use a fuse in virtualized environment.

B

How can we use that in the virtual environment for continuous? Do we need to make.

A

B

Is about what I own FS is the major drive all redhead I see in the community? There are some Chinese colleagues that are already submitting patches. This is the shared file system that lets further virtual machine access. A directory tree.

A

B

Is designed to offer local file system semantics to replace 9p because 9p is do a network file system.

A

B

Now we're tile has restarted and is being developed. It's not finalized yest.

B

What how has been submitted to the community? If you have any opinions, you need to raise about your request for new features or no solutions or any other requests. There need to be added. Actually it includes four parts. The first one is Linux.

B

So you need to add what I all transfer layer above Linux and the second one is key mu. And then, if you refuse and cut a container in 1.7, the modification has already been added.

C

So one block can be.

A

C

You can regard it as.

C

Pages, this is stone. It is stored in pages in the reserve, the regions of the disk where we create metadata.

C

I'll quickly go through this because they sound very details, so this is about the tables actually for fuse. It is different kind of requests and for blob FS. It is different kinds of api's, so you can correspond them together, for example, for lookup. If you are going to look up affair, we have a iteration. We can say that is from the first fell to the next affair and then the cat tribute the open release create delete, read, write, rename flash. So these are the commonly PRS.

C

We have the synchronous, avidians provided for the key new compared to VST. It is totally a synchronous, but the synchronous needs have some limitations means for the underlyings. You do not need to lock, but for the files itself you need to have the lock.

A

C

Are also some details which is about the views corresponding. We can say that it has been embedded in.

A

C

For the general a few requests- and it will also put in Wareham.

A

C

This is about the open or closed operations in the infuse and ASP DK, for example, if owner.

C

Top it is open, then there will be two requestor wires. Look up, the other is open and four look up. We will have the as BFS to iterate to the fair and then open it, so that is kind of preparing resources for close, and it's just the opposite like open. It has two wires release the others forget.

C

The areas related to the SP API.

A

C

This is some of the implementation details with read and write, as well as how you embed the data. So we can say that on the bottom it is POSIX rate. Aladeen has been converted into fuse, read and then to put it into the fuse in header and fuse reading.

A

C

Then, for the fuse in hider and a fuse reading to embed the data into a.

A

C

For the host target, you can say that if it is different there will be a new request. Then it will get the fuse request and.

C

It will call the far ready to to process the request, so can say that this is the first raid and data path on an extension.

A

C

Summarize sums, as I have already mentioned, the blood advance currently can only support limited, fair api's and another big limitations that we can only support a panda right.

A

C

Another thing which is not available: it's tax, it's very useful if you enable Dex or disabled tax. If you do some evaluation measurements, the performance would be quite different, so that is why, for a max, we don't quite realize and then friendly for right. The second point.

A

C

Workload is very friendly, and we also need to think about some amusing scenarios. For example, if you use it in containers, whether it is to write, is friendly or to for grade.

A

C

Red cash feature is very simple.

A

C

Third, one is the big fails.

A

C

Is not friendly to small bells.

A

C

The next step.

C

Some limited scenarios: originally we wanted to use the rocks. Db user scenarios has been changed, so that means we are not limited to the.

A

C

Db, that is why, if going to expand to effuse, we have a lot of.

C

Winners, there is no if it is related to the firm system. So, for example, if you have different requests to approach how to control the different players at this time, you need to add the log.

C

Different kind of fails, you can put it into parallel.

C

It was just realized recently, so our next step planned and we will make it one more either to wait until the commuting is mature. Until the block pass is mature, then we will.

B

Mark a story.

A

C

On the bottom, there are some useful links.

C

You can say that from the second lane that you can get it and for the SP DK reality support. It is on the number one and number three links. So what you need to do.

C

And the benefit of this is that we only need to pay attention to the timing of the block, FS Pro commute kernel and a fuse. We just actively to record our community to do it. So that means we no need to pay too much to the upstream.

C

About our SPD community, currently we've got a different kind of things.

B

C

Have say as Claire black which and then can check some Pascal's and then github Jenkins and the weekly call.

A

C

I knew anyway, it is a open source program. You're welcome to join any questions.

B

About DX I still have some confusion because they so no understanding for the env address and block address needs to have a mapping, and then the mapping will be saved to the page table and then you can access the I mean the physical address, but on this charge I don't see how that does a function. How do you do the mapping? I? Don't understand it fully?

B

Actually, when you look at what I own it has the bar space, it doesn't mean that you need to have physical device for how does simulate M deem you can.

A

B

I can inform my guest, so this part of the memory is persistent and then it will be bypassed in the memory.

A

B

I have to gee sorry I can't find it suppose. I have two G data here. I know.

A

B

I have two G data on the host because if we have requested dahlia and then send it to the guest, so we have a persistent memory with the two G data: it's a device and then on the what I'll text it will be on the hosted memory.

B

So actually, when you have addressed.

C

B

Will bypass that, so it is bypassing this data cache. If you map that, if you have an 2g file space, actually you can see it on the host. Ok, I got it. I have another question about the PMD K I think it's also included in s p, DK p, md, k, NS, p DK are two different things, but normally when we have some it it's together, I think for the SPK in the road map is included. Our road map is saying that we are supporting in PM DK.

B

We can call the API but they're still running their own community. Ok, thank you. I'm still asking about the deck, because you haven't set up about the mapping and I'm a ping SVP I. Can the block do this? Yes, we still do not support random rice, random right. It's not a random right for the to jeepers memory, because you need to map the FD in the key mu.

A

B

Your the host user, Dima I, understand block FS, can only be accessed by demo instead of Cuba for the host solutions. Actually, Kimmo is reducing the whole memory on the host, because chemo is an independent thread and V host and target is also threat. Actually, they can have memory sharing. That's why we require huge TV FS that is Dalia is only the guest memory. Deck Dex is in additional host and.

A

B

Needs mm chain that it doesn't need to change. It also needs to register on a host. I.

B

Remember that we have such a thing you can take. This part of the memory is.

B

Reserved by chemo for 2g space, it will be started together with chemo. That includes 2g memory.

B

That means for the V hosted dinner you're going to operate this. Yes, we can do that, but the key is that what is the reason for you to use pity card? So maybe you want to flush after writing the memory, but our back-end does not support random right. Even when we support the block offers can only do the flush when.

C

B

Support that, because someone has written on the map ready, so it is read-only cash. Yes, we can only have a read-only cash. We can't have anyone to update and right there actually and update that to the car. If we can achieve that, the decks can be achieved if you're doing that, would that be on the V hosted demon right, because the memory belongs to Camille.

B

B

Still wants to asked about Dex I, don't know in the tax Colonel if you used FX. What is the purpose for that is in a must, it's a optional, and we have also market that as unstable. If we don't have these back anything we'll have them ding card, which is for persistent memory. If you don't have that card, when we have M mapped to the DRAM, it can be lost. Actually it will go directly to Peter cache if you're going to write back and flush.

B

Actually, these things in the memory will be flushed to the disk. If you have to G disk space, which is corresponding to the OB space so which side is maintaining the table, if you don't have that, it will go to hit a cache. If you have Dax the back end block device is persistent. That means it will go directly to the block.

B

But if it's a virtual, it's another story, it was just to tell the desk colonel that it is persistent. H wall bypassed the guest, so ultimately it was the still result in the that part of the memory, but actually for the cached data or cache the metadata, maybe you're, connecting that with I meant e. There's no memory block, so the cache heel.

B

Those are you doing that for performance. The purpose is to reduce the petty cash.

A

B

This solution, it's also an host memory and for the guests it's also hosted memory. I don't have physical, persistent device, so the performance should be the same. But when you have a paper cache, you can have some some cash, so you don't need to go to work.

A

B

A

B

Have an Jax device if you're going to write SSD actually is empirical. When you have this device, it's not no longer necessary. I think the reading performance will be worse because I can ask memory, but now we're accessing the disk when you're accessing it's not a faster as petty cash. When you have a physical disk.

A

B

With DRM, of course, the performance will be worse because I've seen some tests for Apache desk on the desk and on the x-axis, sometimes the performance will be worse. Yes, thank you.