#memeconnect #fio
This is a transcription of an interview webcast on SiliconAngle.TV of David Flynn, CEO of hot startup Fusion-io, which makes solid state NAND Flash memory devices OEMed by IBM, HP, and Dell Computer. Interviewing Mr. Flynn are Wikibon Co-Founder and CEO David Vellante and SiliconAngle.com Founder and CEO John Furrier.
DV: In the Cube with us now is the CEO of a hot start-up, David Flynn of Fusion-io. Welcome, David.
JF: The big thing right now, we're hearing from some of the entrepreneurs out there, is pick a market, focus, be disruptive. You're in that right now. Talk about what's going on in your world right now at Fusion-io. You're emerging, you've got some technology. Give us an overview of all that and how VMware fits in.
DF: There's a huge disruption going on in the data center today with the advent of a new type of memory that has 20X the density of RAM and 100X by the time you pack it on a module. Because it doesn't require power and doesn't give off heat. Of course I'm talking about NAND Flash. This is hugely disruptive because you have to say “What would life be like if you didn't have constraints in the amount of capacity of memory and the amount of I/O you have.” It's a really interesting market to be in.
JF: For the folks out there who don't understand all the nuances of storage, you're talking about high-end stuff, but they put this into flash cards for cameras. They put it in, and it works.
DF: That's right. It's the same technology, the same chips that go into smart phones and MP3 players and cameras. But what we're seeing now is that that's being deployed in the enterprise. There's two basic methodologies: One is to have it be a faster disk and pretend to be a disk drive for storage, and the other, which is the route we have pioneered, which was to make it look like a memory device. As a very high density memory it is very valuable.
JF: So we've been covering here at SiliconAngle.tv the convergence – the storage, compute, all converging. Storage has changed. Talk about how you see the storage and cloud changing and the role of storage. It's one big computer? What's your vision; how would you describe that to folks?
DF: it's kind of like in the computer graphics industry, there was a day when you had to have a vertically-integrated platform with vendor lock from say SGI. They could charge $100,000 – even $200,000 – for a single user work-station. That's storage today.
The advent of the graphics card allowed anybody to build a graphics workstation out of commodity parts. So what we're seeing is the storage industry being transformed like the graphics industry was. The agent behind that is the scale-out business model of cloud and high density of flash. The commoditization of the performance layer that used to take so much of an investment.
DV: You mentioned the scale-out. These companies like Twitter and Facebook and the so-called Web 2.0 guys are growing like crazy. And they want essentially building-blocks to scale out. Are you seeing that in your customers?
DF: That's right. And you've already listed some of our largest customers. With these technologies those companies can get 10X the throughput. The benefit of having this high-density memory/storage hybrid is that those databases get 10X the throughput, and their response times are significantly faster. So the scale-out world is being transformed through this new kind of memory.
DV: Wikibon has talked to a number of those customers, and they essentially said that prior to the advent of solutions like yours they were just essentially throwing hardware at the problem. That didn't solve the problem, did it.
DF: And the reason is the servers were just cheap ways to package spindles or package a quanta of RAM. But the CPUs went largely unutilized to the tune of maybe only 10% of the processor being used. The other 90% was sitting idle. So many of these companies bought two-socket motherboards and maybe only put one CPU in one socket and left the other one empty, because if you're not even the full utilization of one why put in two CPUs? It's just cheaper to package your memory in a commodity server package or to package 15 high-performance spindles in a commodity server package than to centralize all that into a big storage array. That's the whole thought of scale-out.
What we do is allow them to fit in the server enough capacity and performance of this I/O memory that they can get full utilization of those processors.
DV: I've been following this market for years, and solid-state disk has been around for a long time. So we've now got this flash coming in from the consumer market, but what else is different now?
DF: What's relevant and the reason why this is exciting today isn't because flash can make a faster disk drive. It always could. You could make a faster disk drive putting RAM with a battery in a disk-drive package. What's interesting is that flash has 20X the bit density per square inch of RAM, and, because it doesn't require power, 100X the density on a module. So it's not flash as a faster disk drive that's made solid state all that important, it's that flash is a higher-density memory device that has really made a big change in the industry.
DV: So the translation is that it's really now cost-effective.
DF: It's very cost-effective. It costs less than RAM per bit and you can pack 100X the density on one module and get 800 Gigs on one card up to 5 Tbytes.
DV: And it's persistent. We've created battery back-ups before, but it's a lot cheaper to do it with flash.
DF: And even if it were not persistent, even if it forgot the data, just as a memory device it would be huge. But the added bonus of the fact that it remembers the data when the power is off is extremely powerful. Now you don't have to wait all that time for all the data to load in. Some of these scale-out properties you were talking about have to stage a database and run traffic against it for several hours before it's ready to go live because they have to warm up all the cache in the RAM. With Fusion-io memory holding the data sets you can go live and it's instantly memory, because the data has been persistent there.
DV: So what does that mean for an application developer say developing cloud applications?
DF: It's a totally different mentality, because things that you thought were impossible to do are now possible. You have to ask yourself, “What would I do if I had a system with T-bytes of memory and didn't have to wait for that to get restored after a power outage but it was holding it all the time.” The fact that this is a hybrid memory/storage device makes for some really interesting designs. You can start combing through massive amounts of data in exhaustive ways that were not feasible before.
JF: David, can you talk about some proof-points in the applications? It's like having RAM in the C-drive back in the old days. What are you seeing in your business that's disruptive?
DF: That's an interesting question because it spans such a large spectrum. We're talking about a fundamental new building-block. So it impacts and will impact everything in the entire data center. In the database world it typically means that a database server can do about 10X the throughput, for the same server. And those queries are answered 30%-40% faster. So it means faster page loads, more throughput per server. So Answers.com retrofitted their MySQL scale-out database tier and saw 9X the throughput per server. What they chose to do was to shrink the database farm four-to-one. So they got a 75% consolidation, and with that remaining one-out-of-four servers they were still getting more than twice the throughput they had before.
JF: So they rearchitected, essentially, their database.
DF: Right. One way to think of what we do is we sell server consolidation, just like VMware, but we do that on the data-intensive end of the spectrum, where it requires such a large dataset or such large access that it wouldn't all fit in RAM, it wouldn't all fit on spindles.
JF: We cover a lot about data on SiliconAngle, and Dave V researches it on Wikibon, and we hear a lot about big data, the Hadoops of the world. We've been teasing out this idea of little data – low latency, fast response time devices.
DF: With virtualization little data becomes big data because you can pack as many workloads as you want on one system. So it's big data at that point because you've jumbled all the I/O together.
JF: So you're fast big data, right?
DF: Right, fast big data. And in combination with VMware what it allows you to do is further increase your consolidation ratio. What limits the number of virtual servers or virtual desktops you can fit on a server is the memory capacity required and the I/O performance required. This fits both of those bills and allows you to completely remove those as constraints.
JF: So what might that mean in a couple of years? We hear about VDI – virtual desktops – and the user experience. But people have iPhones and iPads. So what is that experience going to be? Spectrum is a problem for wireless; everyone wants faster everything now. So explain that enablement; what would be some of the things you would see?
DF: What we're seeing is flash is transforming your end-point devices. So you're carrying around ever more intelligent devices with more capabilities thanks to that very cheap capacity in the NAND flash and ever-increasing processing performance. Where we play is in the back-end infrastructure, what it takes to support those, to run those services in the sky. This makes them much more cost-effective to scale out.
JF: So my iPhone will be more powerful, my iPad will be more powerful?
DF: Those will be more powerful and they will appear infinitely more powerful because they will have fast access to large arrays of servers outfitted with I/O memory that lets them hold huge amounts of data.
JF: It would be like a memory cache that sits in the cloud that serves pages to the edge device, right?
DF: Metaphorically it sits in the cloud and provides a new tier of memory. It doesn't displace disk drives. You'll still use disk drives for the cheap capacity that it gives you. But your active data sets will all be stored in I/O memory. And you'll still use RAM, but you won't have to put as much RAM per server because now you can put in the I/O memory.
DV: So it's a new class of server-based memory?
DF: It is a new class of server-based memory. That is the exact way to put it. The fact that it happens to be non-volatile and it happens to be accessed through the storage interface is really a convenience. Just as a RAM disk sets aside part of your memory to pretend to be a disk drive, our software stack called the Virtual Storage Layer takes I/O memory and expresses it as if it were a storage device. But the way it's physically connected and the way you communicate with it is much more akin to physical memory.
DV: John mentioned VDI. Can you share any metrics or proof-points around VDI?
DF: We set up at the VMware show last fall a demo where we had 512 VDI desktops being served from a single server. The amazing thing is that it's CPU-bound. It's a 48-core server, and we have in essence removed the bottlenecks of the amount of memory required and the amount of I/O. You can log in on an iPad and start interacting with any one of those desktops.
DV: So we've talked about it, but I wonder if you could summarize the differences between the Fusion-io approach, very close to the CPU versus taking SSD and putting it into a disk controller-based assembly.
DF: Two general schools of thought. One is to have it pretend to be a disk drive and sit inside the disk infrastructure, which seems like an easy target for quick deployment. That would be like having flash chips...you know how you carry a USB thumb-drive? You see flash already displaced one form of magnetic media, the floppy drive. Notice, it didn't replace the floppy drive by looking like a floppy disk and going into all those floppy disk readers. There was actually a company that did that, and they didn't last very long, because the USB port was just way to convenient to get more capabilities.
We thought the same thing. Flash is a memory device and should be integrated into the systems as a memory device to be able to truly exploit the additional capabilities. It's more expensive than disk drives, so you don't want to handicap its potential benefit. We should be focusing on having it optimize to get the most benefit from the flash, not having it optimize to look like the old school stuff.
DV: Could you talk about how you could potentially share that resource? Is that potentially on the horizon?
DF: All storage starts local and gets shared out over iSCSI or Fibre Channel through some sort of a director or service. You can take flash memory and share it out through iSCSI or put Microsoft or Open Source on it and put it on a Fibre Channel. So anything that can serve storage over a network can serve I/O memory virtual storage out over a network as well. However, it's best used locally because of the high performance. We have cards that provide the performance of 16 FC4 ports, and it gets very expensive to try to extract that much. So you'll see a tendency, and it's very in line with cloud scale-out, to have the storage local and have it used locally, even if you have a sharing layer of software. For example DataCore, one of our partners, is here showing software that it can create distributed, shared, SAN-like storage from servers that have I/O memory under VMware EFS.
JF: Could we talk a little about Fusion-io the company? What's your status, what's the momentum, revenue.
DF: As you might expect, when you are introducing a new building-block that's this radically different, the question is how do you drive adoption and show that there's traction with it. I'm happy to say that we have OEM wins with all three of the major server vendors – IBM sells our product under the name High-IOPS PCI Express. It's actually designed into their Websphere server appliance and unto their Infosphere appliance, their database appliance. Already built in. HP also has OEMed the product, again sells it as if it were their own, and it's called the I/O Accelerator. That's across their server line. Just a few weeks ago we announced an OEM relationship with Dell who is actually sharing our branding on the product. So it's a Fusion-io branded product. Dell also happens to be an investor in the company.
JF: So are you going to run the table on the server vendors? Dave and I were very excited that you were coming on, by the way, because we think your company's very hot and becoming a huge success. And I/O is a problem. We all want a faster edge device.
DF: We figured that we had better work with the server vendors because what we end up selling is server consolidation. Just as VMware had some of that constructive conflict with “what does it mean to be consolidating servers”, we have that same thing. So we thought it was very important that our products were sold by the server vendors and allow them to make more profit.
JF: It gives them more performance, and still serve their customers and deliver better value.
DF: And the amazing thing is that you might say, “You're going to consolidate servers, that decreases the server spend. But in actuality when you lower the cost per unit of useful computing you make it more cost effective for problems that weren't before, and now you can take business models to market that wouldn't have been viable. So actually it extends the overall IT spend.
DV: It is this thing that David Floyer from Wikibon has talked about a lot – that the compute and storage resources are elastic markets. Lower the price and people will buy more and push harder.
DF: We may talk about consolidation just like VMware, but what it really ultimately means is expanding the overall market.
DV: So you're obviously growing very quickly. Can you talk about head count, revenues, or even just give us some general guidance?
DF: The company is doing extremely well. Steve Wazniack joined the company now going on two years and is extremely active with the company. We're very pleased to have him. Aside from being a celebrity personality, I mean, Dancing with the Stars. Actually that happened just as he was joining up. We thought we would go out on a lark and ask him not just to be an adviser to the company but to actually take a pay check and be an employee and come in on a regular basis as part of the executive team. And he accepted because he is so excited. Solid state has always had a special place in his heart. But then he said, “But I'm going to be unavailable for the next three months because I'm going to be sequestered down in Southern California for this thing called Dancing with the Stars.”
It worked out really well because now people who don't recognize him for being the inventor of the Apple computer recognize him for being that guy on Dancing with the Stars.
DV: So uptake is strong? It's one of those up to the right charts?
DF: The impact it has on business is incredible. We have some customers who are using it under autonomy and Microsoft's fast unstructured text search. That's a growing area of corporate discovery. They went from minute-level response times to comb through every last e-mail, what-have-you, to fractions of a second in the response times. So it can really change the business and the services that they provide. So we're seeing a very rapid uptake.
DV: So the reason I think companies like Fusion-io are so interesting is that not only are you attacking the efficiency play-- consolidation is a big theme. But more importantly, and I think this is true for VMware as well, helping IT become a value producer, whether it's a revenue generator or a new-business enabler. From your standpoint it's enabling orders-of-magnitude greater application performance.
DF: That's a good way to look at it. And we segment it too. There's folks who want to use our product as a way to reduce cost and improve efficiency. The more forward-looking guys are those who look at the creative ways to – now that this is possible and cost-effective what creative things can I do with it that I couldn't before? It means thinking outside the box.
DV: More tactically ,what worries you as a CEO? What keeps you up at night?
DF: My biggest worries these days are around supply chain and building things fast enough. It's a problem that we always said, “Hey, it would be great to have.” Well, it's still a problem.
JF: Just a final question, advice to other entrepreneurs. You've built a great growing business. What do you tell other entrepreneurs out there?
DF: Run with your dream. Analyze it well, make sure you're targeting a good market, and just trust your gut on stuff. And be a little bit of a non-conformist. What's really worked well with us is everybody else is trying to conform to those same drive bays and storage controllers, and we said that all needs to be rethought. So being ambitious enough to take on daunting tasks.
