Original URL: http://www.theregister.co.uk/2011/06/25/violin_memory_billion_dollar_platform/

Violin tunes up for billion dollar flash gig

Playing the big revenue game

By Chris Mellor

Posted in Storage, 25th June 2011 12:00 GMT

Violin Memory CEO Don Basile reckons Violin can repeat what Connor Peripherals did in disk drives, Cisco in networking, and NetApp in file storage; become a billion dollar a year revenue platform company.

He'll do it with flash array technology because spinning disk is going to become rust, failing to keep up with the I/O demands of voracious servers. The enterprise IT buying market is beginning to share this perception because the biggest mouth in the industry backs it - Larry Ellison.

Larry Ellison's turtle speech

Basile thinks Oracle boss Larry Ellison - "the public spokesperson for the entire IT industry"- altered enterprise CIO thinking with his Turtle speech. Ellison introduced Oracle's SPARC-based SuperCluster in December 2010 and said it was a cheetah racing along at more than 30 million database transactions a minute.

Turtle

A SuperCluster system, configured for a TPC-C benchmark, comprised 1,728 processor cores in 108 T3 SPARC CPUs. It had 13TB of DRAM and, something that excites Basile, 246TB of flash memory, backed up by 1.37PB of HDD array storage.

Ellison said IBM"s Power 7-based systems were like race horses, running at just 10 million transactions a minute, while HP's Itanium-based Superdome clusters were like turtles, managing just four million transactions/minute.

Basile said: "Larry is telling people to use flash … That's the fundamental shift in the industry. … Customers know their competitors will adopt the technology. Will they be first, second or last in their industry to do so? … It will happen and happen relatively quickly. It's not just speed; its the lowest cost of data base transaction in history. [Flash] is faster and cheaper on the exact same software. It's a no-brainer."

"We want to achieve a billion dollars in less than four years. It's possible."

The Ellison message, for Basile, was that the only way to keep the processor cores busy is with a great big slug of flash memory in front of the spinning disk arrays. It responds to read I/O requests in microseconds instead of the array's milliseconds, and does it without a high-end disk array's physical size, electricity needs and controller complexity. There is no way that spinning disk arrays can match NAND array performance or its far superior cost per transaction, and that's with today's 2-bit multi-level cell (MLC) NAND chips.

This is a strand of thought we have picked out from recent benchmark scores such as a TMS RamSan-630 SPC-1 result, about which we said:

There is surely no point in IBM – or any other vendor – trying to beat TMS on the SPC-1 benchmark with anything other than its own flash-based system.

Basile is convinced that CIOs now understand that they have to have database and other transaction processing systems in the future which use clustered processors and back-end bulk data tub drive arrays with front-end flash arrays to hold the dynamic data, the working set data. This is a theme coming out of the high-performance computing and big data area also, witness an Intel presentation at ISC in Hamburg.

However, it is the case that general and widespread use of front-end flash arrays won't happen unless flash costs go down. Basile thinks that stopper is going to be cancelled out and a tipping point reached as flash arrays cross the chasm into mainstream deployment. The technology that will make this happen is 3-bit MLC.

Three bit MLC

With three bits per cell 3-bit MLC NAND should be 50 per cent cheaper than 2-bit MLC per GB. But there is a drawback; 3-bit MLC has a shorter working life than 2-bit MLC and flash controllers using it have to get around this with more complicated wear-levelling schemes, error-checking-and-correction (ECC) logic and so forth. It is non-trivial.

Violin seems confident it can pull this off and deliver 3-bit MLC arrays: "We see 3-bit MLC coming in, early next year, in volume by summer. The amount of error correction required is greater than 2-bit… needing 100-bit, even multi-100-bit ECC corrections. To do this at speed is not easy."

Violin array flash cards

Flash cards in Violin Memory array enclosure.

He is confident it can be done and says: "The economic impact is that we'll be down to $2/GB level pricing."

Today's flash pricing sees Violin list pricing MLC at the $20/GB area and transacting orders in the mid-single digits. SLC flash product lists $30/GB and transactions take place in the teens. If Violin can transact 3-bit MLC flash product orders at $2/GB the financial logic is compelling - as long as volume sales result.

If this happens and volume flash array, and SSD, sales take off then ensuring NAND chip supplies will become paramount.

NAND chip supply security

Roughly speaking Samsung and Toshiba have 80-90 per cent of NAND flash fabrication volume with IMFT (Intel Micron Flash Technologies) having the rest. We are NAND supply-constrained. Fabs cost $10bn plus and there aren't enough of them even to begin replacing the spinning disk industry even if NAND was as affordable as disk.

Basile said: "We made our deal to make Toshiba our largest investor and strategic partner - so we will get our NAND. Steve Jobs will get his NAND and we will too."

If NAND prices do come down with 3-bit MLC then demand will jump and supply constraints occur. If that happens then having security of supply for a NAND-based product supplier will enable them to continue uninterrupted production and not suffer price spikes. SSD and flash array suppliers buying NAND on the spot market will have to pay higher prices and be disadvantaged.

What existing NAND supply arrangements exist?

Toshiba has invested in Violin Memory and has a strategic partner relationship. SanDisk partners Toshiba in the fab business and has bought Pliant to get into the enterprise SSD business.

Basile said: "SanDisk bought Pliant because the enterprise space is really big. It became a new entrant in a growth market. They should probably buy another couple of companies if they want to be taken seriously, getting software and network expertise."

Samsung has invested in Fusion-io and has a relationship with Seagate for SSDs.

Micron makes PCIe format flash cards and SSDs. Intel makes SSDs, has a relationship with Hitachi GST to make additional SSD product, and will likely enter the PCIe flash business. Western Digital is buying Hitachi GST and should inherit the Intel relationship.

NAND product suppliers such as OCZ, STEC, and Texas Memory Systems, buy on the NAND spot market.

Basile said: "We made our deal to make Toshiba our largest investor and strategic partner - so we will get our NAND. Steve Jobs will get his NAND and we will too."

Don_Basile

Violin Memory CEO Don Basile

"In a high-demand market prices spike up, not down, and it's complex to hedge… NAND is the fuel of the industry. if prices go up and you have no partner, you have a spike risk. Steve Jobs has his NAND fuel; he's hedged. You take a 20 per cent NAND price hit and if you're OCZ you're in negative margins."

If there are supply shortages then spot market NAND buyers may not be able to fulfill contracted delivery obligations. Basile thinks this means: "a lot of risk to the business model … we've mitigated the risk."

Over time, as NAND dies get smaller they are less and less interchangeable. Merchant NAND buyers will then find harder to leverage the spot market.

So suppliers like OCZ, STEC and TMS share a vulnerability, and this is "a huge risk for them" according to Basile.

Three directions for the enterprise flash industry

Basile reckons the enterprise flash industry can go three general ways. One option is that the existing storage suppliers "leave it alone and new giants emerge, like Cisco did in networking".

A second possibility is that existing NAND "suppliers in the industry come and buy up-stack." He likens this to what Intel has done, moving up stack from processor chips to chipsets and on to motherboard design: "Fundamentally most server IP today is all Intel. It's especially so with Sandy Bridge."

Basile pointed out: 'Intel has always failed on the system side. It's fundamentally a semi-conductor side company." So it's unlikely to become a server company.

The third way is for storage-supplying systems companies to buy down-stack: "But storage guys never do vertically integrate downwards." EMC and NetApp, for example, haven't bought disk drive suppliers. The systems mindset is completely different from the component mindset.

Basile hasn't seen much innovation inside storage array suppliers: "The main storage vendors took out innovative newcomers, like 3PAR and Compellent. … They didn't innovate in-house. They bought IP instead." It seems to us, after talking to Basile and Stephen O'Donnell, a Violin Industry Advisory Board member, that they think the main storage suppliers are unlikely to buy a memory array competitor because its technology competes with their own.

Storage suppliers like EMC, Dell and NetApp are putting array-controlled flash in application server arrays though. Why? "They can't eat their own baby food," is what O'Donnell thinks. They want their spinning disks to keep feeding data to app servers and not be relegated to tier 2 storage duties, feeding silicon storage instead.

Violin's strategy

The Violin strategy is to focus on building network-accessible, low-latency, high IOPS silicon storage arrays, riding the flash technology development curve, particularly 3-bit MLC, and sell to the global 2000 enterprises through a combination of OEM and distribution/reseller channels.

Violin

Basile said: "We focus on delivering highest I/O at lowest cost with the most reliability. We should emerge as one of the winners. It's not an uphill sell, it's a downhill sell."

El Reg understands that a likely next-generation product from Violin may arrive in the fall, in autumn, and do two million IOPS from a 2-3U enclosure.

For a flash storage company to emerge and stand alone it has to achieve escape velocity from the main server system and storage array vendors. It has to do what Connor Peripherals did in disk drives, NetApp in filers and Cisco in networking. The technology has to be difficult to emulate, the technology disruptive to existing suppliers and hence unattractive to existing product line bosses, and the marketing and channel strategy effective.

Basile says that Violin's "billion dollar revenue will come from channels, both OEM and resellers. … Connor went from zero to a billion in four years. … There needs to be a distribution channel with an OEM one."

"We want to achieve a billion dollars in less than 4 years. It's possible."

"We're focussed on the global 2000 and verticals; Internet. financial services, and government. All of them are moving into data analytics and computing on that data faster and faster. They collectively spend $1.3 trillion a year on IT. It is not that hard from that point of view to extract a billion dollars of revenue."

He says that there is no way spinning drives can service their compute needs in the future; "Latency is the critical factor."

Yes, Violin is a start-up and relatively unknown, but "we have a partner next to us - HP - taking out the brand risk".

The opportunity is increased because "Every disk array in existence will go end-of-life in 36 months and need replacing. Your storage comes off cycle/maintenance in that time frame." They have to be replaced.

Violin, Juniper and IBM

Violin is working with Juniper, and has an investment from that company, because there is a good marketing fit between their two strategies.

Basile characterises Juniper's QFabric as a network that can deliver deterministically low-latency, non-blocking traffic, and can scale. You can put a flash array, clustered servers and the QFabric together and arrive at an optimal data centre scale architecture, with linear scalability of compute and I/O.

He says that Violin's memory array is fundamentally a networking device needing a network to connect to servers. This could be HP's vConnect backplane in a bladed server system or QFabric at a data centre level.

Juniper will ship QFabric in the fall, with Violin Memory arrays to drive traffic through the network.

Basile said Violin is also working with IBM and its General Parallel File System (GPFS): "We will make GPFS extremely fast and high performance, modifying 10 billion files in 42 minutes. … With two 2-racks it will modify a trillion files in 3hours … sharing a single global namespace.We worked with GPFS for over a year to make this possible and achieve linear scalability. Both meta data and bulk data are stored on Violin's arrays with GPFS backing off trash to other storage. It will completely change the way people build storage pools.

Back to the turtle

Basile said: "Ellison's Turtle speech changes everything. How do you not buy memory arrays to run Oracle solutions?" For both server system and storage array suppliers the questions becomes "How do we turtles transform ourselves into Oracle-chasing cheetahs?"

Larry Ellison

Oracle CEO Larry Ellison

HP and IBM are doing it with Violin. Fusion-io no doubt has its main server partnerships working to the same end. But which way will Tucci and Gelsinger at EMC jump? What will Georgens at NetApp do? Will these two storage bellwether companies collude in the destructive downgrades of their own HDD array products by investing in memory array technology?

Imagine a Violin-assisted integrated HP stack competing against a vBlock with no front-end flash array; the HP system will likely blow the vBlock away on price/performance at a transaction level - remember the RamSan-630 SPC-1 benchmark mentioned above?

Imagine an IBM SONAS flash-assisted scale-out filer competing against a NetApp filer or EMC VNX or Isilon system. Big Blue will walk all over EMC and NetApp in that instance.

The import of Violin CEO Don Basile's thoughts, if they are correct, is that EMC, NetApp, and every other enterprise storage array supplier has to embrace solid state array technology and see their HDD array business become less and less important.

They will see this outcome whether they embrace memory array technology or not. If they do embrace it then they become players in a high-growth market and will still have growing businesses.

If they don't, they will be legacy players in what will become a low-growth market, possibly a no-growth market - if Basile is correct. That is a real big "if".

Is he making sweet music, then, or just making screeching noises? That's what Gelsinger, Tucci, Georgens and other server system and storage arrays bosses have to decide. ®