For Fujitsu Siemens Computers, solid state disks are a datacentre issue. In our report on its Augsberg VisitIT conferene in October we wrote: "The company's chief technical officer, Joseph Reger, thinks that the performance and economics of flash memory-based solid state disks (SSD) would result in substantial datacentre take-up as soon as 2010."
This was an intriguing idea, running contrary to the prevailing view that SSDs would appear in ultra-mobile devices and laptops before any datacentre penetration. We asked for an opportunity to talk to Joseph Reger and find out more about the reasons for this prediction. Here is the result:
Techworld: Could you discuss the cost per gigabyte trends affecting SSDs?
JR: It's a bit of a complicated picture. As I said at VisitIT the cost difference in all respects will be gone by 2010. In almost all cases the HDD and SSD cost/GB price curves will have crossed. The first curves that cross will be in enterprise storage in 2009 or so; the first to be affected will be Fibre Channel (FC) drives, then serial-attached SCSI (SAS), and then serial-ATA (SATA). By the end of the decade hard drives will have difficulty competing.
There are economies of scale and the speed of development of semi-conductor technology, Moore's Law. We're seeing increases of two times the density and capacity every 18 months or 2 years, faster than hard drive developments.
Flash is also moving from single layer to multi-layer in the cell, increasing capacity. A current problem is that flash has an upper limit of re-writes. What I expect to happen is that with increased capacity there is more opportunity for wear-levelling. The algorithms will also improve and so the problem will become less.
Beyond 2010 I expect SSDs to compete with hard drives in all areas.
Everybody was expecting SSDs first in laptops. I expect a volume market first - enterprise SSD use. Laptop hard drives have a low cost. They will be harder to attack. On the enterprise side, in the high IOPS transaction area, we don't even have to hit the same price. We just have to hit a price point. In this random IOPS area even for a little bit of a premium SSDs can show very impressive performance. Today, SSDs can deliver 200 times faster performance than hard drives.
Techworld: How will SSDs be used in datacentres?
JR: We can think about meaningful mixtures of old and new technology. Database index files can be stored on SSD with bulk data on hard drives. This is a hybrid mix. There are already some hard drive arrays with gigabytes of flash cache. That will be a very interesting enterprise play, an interesting discussion on how to do this right. We'll have to re-think how RAID is implemented. It can be done at the hardware or the chip level. But RAID at the controller level does have merits. We might have intelligent controllers to put data appropriately on SSD or HDD.
My personal opinion is that the traditional array vendors will do this. It's a very interesting process right now. Big vendors change their message every day.
Seagate once said it (SSD) will never happen. Now Seagate is saying it will do it itself. The traditional manufacturers are changing their minds.
The digital photography parallel
It could be a similar process to what happened in the digital camera industry. Digital cameras were a disruptive change for the film-based camera industry. Some companies big in film cameras saw it in time, reacted, and made a digital camera investment, i.e. Canon.
Some came late but did manage, Nikon. Others did not and disappeared, i.e. Minolta.
Others were not camera vendors at all but did have the digital technology and became successful, i.e. Casio.
Samsung and SANdisk could become SSD array vendors.
It could well be the storage array contributors who succeed because they do the controller right. They could conceivably have such an influence on the way the controller works that they could dominate, even though they don't make solid state drives.
The wear-levelling algorithm is key. It will be helped by having more SSD real estate.
Techworld: What about RAID and SSDs?
JR: RAID in its current form is not needed in SSD arrays. Some of the functionality could be done at the chip level.
Techworld: Could you discuss how the SSD array adoption might occur?
JR: I think the way into this technology will be through replacement. We have applications, the operating system, drivers and then the array backend with a hard drive controller and hard drives. It's easier to just change the back end, to swap out the hard drive controller and drives for an SSD equivalent. The front end just will not change.
I'm a strong believer that, typically, with disruptive change all the assumptions made by traditional vendors are revisited. Some suppliers coming in to respond will take a very hard look to find differentiation and may well change the architecture.
Techworld: How will SSDs and storage virtualisation interact?
JR: I think the SSD introduction will be greatly facilitated by virtualisation. It can facilitate the swapping in of the new technology as a replacement. It's something we are looking at.
Even if the technology is interesting there is still this aspect of risk while introducing it. We learnt this with our CentricStor virtual tape product. It's the functionality and performance of the virtual system that sticks in peoples' minds.
In the beginning we'll have mixed environments and that's great for virtualisation. We'll hide the HDDs and SSDs in effect. The mixed operation is just great.
Techworld: Any view on thin provisioning and SSD arrays?
JR: Thin provisioning does help for the capacity market. It lets you get away with less hardware. It's important here because capacity prices for flash are not that great (and you can defer purchases and benefit from decreasing flash prices.)
Techworld: What about power consumption?
JR: One of my favourite reasons for liking SSDs is the power consumption. It is about a fifth of a hard drive array for the same capacity, in some cases even more.
In the enterprise high-performance arrays take a lot of power. There it could be even more spectacular; I would expect just ten percent. Here, for extra performance, some people are using just the edges of the disk; just half of the capacity is being used and so over-provisioned for speed. This can go. It gives you another power consumption drop.
With less power in there is less need for cooling because of that. SSDs also need much less floor space altogether. This will help too.
The case made in favour of SSD penetration of the datacentre is persuasive. It's not just the forthcoming crossover of the flash SSD and HDD price/GB curves. Performance-hungry applications using FC drives will find very much faster SSD arrays attractive. Increased SSD cell capacity through multi-layering will enable many more bits to be used and so spread out the write load, reducing the impact of SSD's limited write cycle numbers. As well as such capacity increases mitigating the problem improved wear-levelling algorithms should mitigate it too.
The power and cooling implications of SSD adoption could be spectacular with the replacement of drive edge-biased FC arrays resulting in a ninety percent drop in the power draw, reduced floor space and an overall lower cooling power budget.
The existing storage virtualisation products will make it relatively easy to swap HDD arrays out and SSD arrays in quite transparently.
The digital camera industry analogy suggests that not all HDD array vendors may survive: EMC, HDS, IBM, LSI, Xyratex, Dot Hill. It suggests that flash technology vendors such as BitMicro, Micron, STEC, Samsung and SANdisk may have a role and, also, controller vendors who make best use of flash may have a strong role. DRAM SSD vendors such as TMS, already producing a DRAM-cached flash SSD, will also be players as may drive vendors, such as Seagate, who embrace SSDs.
It is going to be a very exciting few years in the enterprise array space because flash-based SSDs are a true disruptive technology.