Host Enrico Signoretti speak with Martin Fink and Chris Bergey of Western Digital about zoned storage and what that means for the data storage industry.
Martin Fink Martin Fink joined Western Digital as Executive Vice President and Chief Technology Officer in January 2017. In this role, Fink leads the company’s technology innovation agenda as Western Digital continues to transform its business and expand its capabilities and technology portfolio.
In his 30-year career at Hewlett-Packard Company and Hewlett-Packard Enterprise Company, Fink worked in a wide range of roles, including General Manager of HP Cloud and Senior Vice President and General Manager of Business Critical Systems and Converged Application Systems.
Chris Bergey Christopher Bergey is responsible for developing and driving Western Digital’s embedded storage solution strategies in the mobile and connected market segments, including smartphones and tablets, automotive, industrial, connected home and other connected, “Internet of Things” environments.
Enrico Signoretti: Welcome, everybody, to a new episode of Voices in Data Storage, brought to you by Gigaom. I'm your host, Enrico Signoretti, and today we will talk about a new industry initiative around data storage. This initiative was launched a few weeks ago by Western Digital, and it’s called Zoned Storage. Helping me-going through all the details about Zoned Storage, today I have Martin Fink, CTO of Western Digital, and Chris Bergey, Senior VP of Devices. Hi guys, how are you today?
Martin Fink: I'm doing great! Thanks, Enrico, hope you are doing great as well.
So, Western Digital, why don't we start with a little bit of backend about your company. I know that everybody that works in data storage should know about it, but just to start with a little bit of information about Western Digital and why we are here today.
Chris Bergey: Sure, Enrico, this is Chris. So, Western Digital is a global leader in storage. We're actually kind of the coming together of three large storage companies between traditional Western Digital, Hitachi Global Storage, and then most recently SanDisk, which brought the NAND portfolio to part of Western Digital. So we're offering a full array of products from traditional hard drives, to capacity enterprise drives, as well as flash products for both end points and data centers.
Fantastic. So we're talking about Zoned Storage. Before going into the details about what it is and what are the benefits of Zoned Storage, why don't we start with the needs. Why did you start thinking about Zoned Storage?
Martin Fink: So, this is Martin. Let me take a little bit of a crack at that, because I think there's a couple of angles to it. One, quite frankly, is customer pull. And probably the most important is when our customers are beginning initiatives on their own to try and maximize the amount of information they can store at the lowest possible cost. And they're essentially telling the industry that the devices and capabilities that they're getting from the industry aren't quite meeting their needs. We have to listen to that. So that was definitely a significant part of beginning the Zoned Storage initiative- is feedback from the customer.
The next part is that Western Digital is in this unique position of being both in the hard drive business and in the NAND flash business. And as a consequence of really being deeply engaged in both, we essentially saw an opportunity where these technologies can come together, especially at the software level, to really bring a new set of capabilities where our customers can leverage a singular software stack to be able to talk to both the hard drive and the flash in this new Zoned Storage mechanism that maximizes the amount of data they could store, at the lowest cost per bit, with various performance options they can choose from. So, it's really this combination of the customer pull and this unique capability around being exposed, both the hard drive space and the NAND flash space, that really kick started this initiative for us.
So, we're talking about large customers here, maybe hyperscalers, ok? They want to store huge amounts of data in the cheapest possible way. And you started to think about how you can leverage both NVMe flash, down to the biggest of the hard drives, right? And put them together in an end-to-end architecture that can help them to achieve also performance in this picture, right?
Martin Fink: Yes, so the - in the hard drive space, our hyperscaler customers, as you talk about, some of them had started to use a capability known as SMR, or shingled magnetic recording, in order to be able to get more utilization and be able to store more data on the devices. And this SMR, or shingled magnetic recording capability, required them to do - to use a special software layer that's now part of the Linux kernel in order to be able to leverage the capabilities of shingled magnetic recording. And then - again, because we were also in the NAND flash business, we saw a very similar capability evolving on the flash side, and an extension of the NVMe standard called Zoned Named Spaces, or ZNS. And we saw the opportunity to essentially use the exact same software stack, or evolve the same software stack, that we'd been using for SMR on the hard drive side for those Zoned Named Spaces flash SSDs. And that's the part that also really started to get the attention of our customers.
I might be confused now because I totally agree with you. Ok, SMR needs something that organizes information before they reach the drive, ok? Because that is around it. But, the industry taught me that flash is totally different story, ok? Why we are starting - talking about this also for flash devices?
Martin Fink: Because that same software that stages all of the device rights up on the host is the capability that the flash device is used for ZNS. And as a result, moving all of that work of staging the rights for video data, IoT data, that kind of data on the host, what it allows us to do is create a flash device with a significant reduction in D-RAM, and a significant reduction in overprovisioning. So it's those capabilities that are now very, very similar to SMR on the hard drive side, this notion of staging the data on the host before we write it to the device.
This also means that with the advent of 3D QLC NAND, we will have an improvement also in performance for these devices.
Martin Fink: Yeah, Enrico, I think you're not confused, actually. The reason that an SSD is able to handle random writes, or as you kind of mentioned, doesn't have the limitations of SMR, is because we actually, basically build a translation layer between the NAND media and the host. And that translation layer was required to make it look like a hard drive, which was the initial application for SSDs where we'd build up our drives. We wanted to drop in maybe faster alternative, which was an SSD.
But as mentioned, that actually causes quite a bit of cost and also things like over visioning and really inefficiencies. Because if you go all the way to the NAND media, we actually have to write to that fairly sequentially as well because of the way that NAND needs to be raised at the block level. And so there's a lot of management that actually occurs by the SSD in actually programming the bits and erasing the bits. And so ,what we're doing with ZNS is actually simplifying that, much like we do on SMR. That does leverage sequentialization of the data while still giving you all the advantage of random reperformance.
And you're right. And then that also prepares us to take advantage of new technologies, like QLC, where maybe there's additional tradeoffs on write performance, but you still get very high-performance reads. So I think the similarities are a lot closer than people realize. And this is the synergy that we see between our HTD and our NAND divisions.
But does it also mean that in the future, we will see more optimized device? I mean, we need part of this stack, ok? So once it was the emulation of the drive. Now that you have NVMe, you are moving some of this intelligence to the host that organizes information. Does it mean that we will have faster devices or even more efficient devices from this point of view and probably, also, from the economic perspective, cheaper devices, cost effective?
Martin Fink: Okay, so let me try to see if I can segment this clearly. The Zoned Storage Initiative, SMR, and ZNS drives are really targeted at a capacity workload. So where you need a significant amount of capacity and you need reasonably good read performance. That is a very distinct segment from we'll call it the high-performance space where a conventional enterprise SSD, which is another part of our portfolio. So, I think it would really help the listeners to kind of really keep in their mind, separate, that the Zoned Storage Initiative is about low cost, high capacity, with reasonable read performance. So our goal in Zoned Storage is not to achieve the high performance, high throughput, everything about NVMe, and fabrics, and all of those things. We have a separate category of products for that. So that's sort of a simple way to just keep it separate in your mind.
Ok, very good. So, but that doesn't change much. I mean, you will get - so by taking advantage of the Zoned Storage techniques, you will be able to get cheaper devices, or even more efficient devices.
Martin Fink: Yeah, you will basically get the lowest TCO or - and the lowest cost per bit for a given performance point using Zoned Storage.
Chris Bergey: Enrico, I would highlight that that may not even be the most compelling part of ZNS. One of the most compelling parts is actually the idea of being able to add this intelligence of data groups or data streams. So for example, a host may know that let's say a few gigabytes of data are going to be placed on the drive. That it knows is a high probability that they would want to access that all at the same time or erase all of that at the same time.
Well, if they can actually provide to drive that intelligence that that data set should be placed locally on the drive in the physical locations that are adjacent, that actually has the opportunity to drastically not just impact the performance, but impact endurance, reliability, and all those kinds of things. Because in today's drives, we've got a fixed hour then that's maybe making some assumptions on what data should go together. But, it doesn't have the capabilities at the host level that's provided. So yes, we're leveraging costs. Yes, we're leveraging getting to the highest capacity as possible. But there's a tremendous amount of value at the system level with this type of solution and with this type of intelligence.
Ok, do you expect that alongside hyperscaled you will have support for this initiative from large enterprises or any other kind of end users as well?
Chris Bergey: So I believe the answer is yes, Enrico. What you saw even if you look at the hyperscale transition, right? They, when they started, they were building the biggest, baddest servers they could build, again, using traditional architectures. And only say ten years ago did they start pivoting and saying, you know what, we could do this better because of our size and scale. And they started looking at things like distributed computing. They started using different types of file system capabilities, different type of redundancies at different levels. So that- those techniques have been driven by the hyperscalers.
And you do see these technologies trickle down into the more traditional markets and the more traditional applications. And so that's really what this initiative is attempting to do is to share some of these product concepts, product ideas with the larger ecosystem to drive a wider adoption. Because it is a little different and scary for people where it's definitely not just redoing or doing what you were doing previously. Creating file systems, creating systems that are able to sequentialize data is not simplistic.
And that's why Martin said that it's not for all parts of the data center. You are going to have workflows that are going to be close to the processor that you will want to keep quite random. But that's not actually the largest portion of the data growth today, which is able to take advantage of many of these new techniques.
Yes, and also with the growth aside that we are seeing in SSD devices on one side, and also these SMR drives. I think your roadmap talks about 20, 25 terabyte drives in less than two years from now, right? So it's important to organize information better in the single device and in the entire correct?
Chris Bergey: Yeah, so we - one of the things that we announced a few weeks ago was that we were demonstrating our 20-terabyte device that we plan to bring to market next year. And that's leveraging SMR technology. And we do expect that actually 50% of our bits that we ship on our drives will be SMR-based by 2023. So that's not too far from now. But what we also see is the capabilities of the technologies that when people re architect their systems to take advantage of SMRR drives or QLC ZNS SSDs, that the technology barrier, we're able to really achieve another step function of bit growth and cost per bit. So that's really the big payoff over time. These architectures take time for implementation, but there's a significant payoff over time.
And one question that comes up to my mind right now is okay, so we are planning for the future, so Zoned Storage, architectures, and okay. What happens if I have a mix of old hard drives and new hard drives in the same data center? Can I take advantage of the same libraries, of the same methods to brighten all the drives as well?
Martin Fink: That's probably not going to be the best way to do this because one is you must have a SMR hard drive to take advantage of it. And then using a high-performance flash drive is probably not the most sort of effective use of your capital resources. So I think it is a - it would be better for customers as they look at their data center to deploy the new workloads that have a high sequential write profile onto new SMR and ZNS drives. And then perhaps redeploy their existing fleet of devices to applications that can leverage the performance of those devices.
I see, I see. And what about the developers? Where they can find information about Zoned Storage?
Chris Bergey: Enrico, we've actually started, as part of this initiative, zonedstorage.io, so that's Z-O-N-E-D-S-T-O-R-A-G-E- dot I-O. And this site is intended to really help developers understand what work has been done, what's available on the - in the open source community. Some of the partners and supporters of this effort and of the general industry moved to Zoned Storage. So, that's where I would recommend developers start. And we are very open to additional contributions, additional concepts and ideas.
But that's the place where developers should start. And we're really trying to support and encourage a large, open community in support of this effort. Cause as I mentioned, it does take some time. It is a significant change. But as kind of Martin mentioned, there is just such an onslaught of data coming. Much of it is highly sequentialized from an IoT or video point of view. But there are obviously opportunities to even look at traditional workloads and how you sequentialize those. So it is a - it's a difficult task, but it's a necessary one. And we see success happening. And we're just looking for a larger ecosystem of support. And we think it's essential for the industry to grow and to support the data growth that's really upon us today.
So, you are looking also for partners, then. I mean and which kind of partners, in that case?
Chris Bergey: Well I think that Western Digital has established itself as a very open company. And we're very supportive of open initiatives. Martin has been a big part of that. But, I think that there's a lot of requirements at the database level, at the OS level, at the - as well as at the hardware level of support. So it's an open initiative right now.
Martin Fink: Yeah so, that's a great point. So the- everything that we're doing related to Zoned Storage is an open source initiative. And so we have been doing most of the work on the Linux kernel level. But then as we start moving into file systems, databases, and more of the middleware space, that's where getting more help from the developer ecosystem in each of those open source projects would be great. It would be phenomenal, right? It would be allowing all of those open source projects to take advantage of that increased capacity at a lower cost. And then, but we also want - and the reason why all of this is also open source, we want the end user, the customers who have the actual data that needs to be stored, to also have access to all of those open APIs so that they can prepare their data and get ready for all of the various devices for Zoned Storage. So that's kind of like the layers and the stack as it were.
So - do you have any partner or other industry vendors supporting this initiative already?
Martin Fink: So this is a very early initiative. And one of the things to keep in mind is that we're building upon a leadership of SMR that we've built over the past number of years. So, an example customer that's been an early adopter of SMR is Dropbox. Now, we've got other partners in the ecosystem like Broadcom, SUSE, ATTO, and others that are also doing that. On the flash side, the ZNS standard will be ratified later this year. And so that's a little bit - it's a little bit early on the flash side. But we definitely expect the - because we've had a lot of customer pull, and we have been speaking to customers. We expect more and more customers to get on board to the Zoned Storage initiative as a result of being able to use that one software ecosystem for both their SMR fleet and their Zoned Name Space drive fleet.
Chris Bergey: Yeah Enrico, the other thing I might add is beyond the NVMe working group, which has a lot of interest as Martin mentioned, you can actually also look at the Open Compute Project. And you can see there's quite a few submissions and discussions around ZNS. And there's a lot of good information there as well.
So listening to you, that the initiatives looks very interesting for hyperscaler, people that already adopted SMR. And maybe there is an interest also from other organization, like Open Computer Project and things like that. Am I right?
Chris Bergey: Yes, Enrico, you're right. I believe if you go back to the Open Compute Summit that occurred this spring, you can find quite a bit of ZNS discussion in some of the presentations and working groups. We also plan to be a part of the Amsterdam OCP event coming up this fall where we're going to do more discussions around ZNS and our Zoned Storage initiative.
That was a fantastic introduction, I think, about Zoned Storage. Thank you very much, guys. To wrap up this episode, I want to give, again, the link for where you can start to learn more about Zoned Storage, which is zonedstorage.io. And if you want to - you can start from there to get the technical information, open source libraries too, and our sources for data center architects. Thank you, again, Martin and Chris, for your time today. And bye-bye.
Chris Bergey: Thank you, Enrico.
Martin Fink: Thank you, Enrico.