The death of disks

by Robin Harris on Tuesday, 25 June, 2013

A forecast says PC shipments with disk drives will drop by a third between now and 2017. IBM is pushing the all-flash datacenter. SSD start ups are claiming that flash is really as cheap as disk with much better performance. Is it the beginning of the end for disk drives?

Disk is the new tape and not in a good way
Everywhere you turn pundits – and vendors – are predicting bad things for disk drives. But are they correct?

SSDs cost many times what disks cost. There’s no magic here: the NAND flash chips simply replace the heads and platters of disk drives with much more expensive transistors. Where do these numbers come from?

From several assumptions. In no particular order these include:

  • Reduced power and footprint costs that come with the compact and cool running SSDs.
  • Reduced management time and effort required to optimize performance.
  • In-line compression and de-duplication reduce needed capacity.
  • Two-thirds of unused enterprise disk capacity wouldn’t be needed due to the superior performance of flash technology.

Let’s look at each of these in turn.

  • The power savings are real. But in the average data center power is a single digit percentage cost. And there are many other consumers of power that make the savings even more marginal.
  • Floorspace savings are also real but unless you’re building a new data center or are out of space in your current data center, it is unlikely that you’ll reduce the size of your data center enough to claim much advantage from flash’s high density.
  • If your data center is already full, which is not uncommon in densely populated areas, you can save big. In parts of New Jersey co-location space rents for four times that of high-end office space in New York City. There flash can save you some real money and, not coincidentally, make you real money in the financial markets.
  • It is also easier to achieve high performance with flash arrays than it is with disk arrays. In fact, system and storage admins often need to unlearn techniques that were developed to optimize disk performance when they start using flash. But unless you’re going to fire your database and storage admins you are unlikely to actually save any money by moving to flash. Those people will move on to other work.

Unused capacity
The argument about unused capacity comes from the practice of short stroking expensive 15k Fiber Channel and SAS disks to drive IOPS. But that’s not the only reason for unused capacity. Other reasons include:

  • It takes so long to order and standup new storage that everyone is forced to over-configure or risk running out at critical times.
  • Enterprise storage is typically purchased on a project basis. When a new application is brought up it often gets its own physical infrastructure. Since the growth rate for the application isn’t understood, the natural inclination is to over-configure.
  • Storage and application admins typically have only a hazy idea of performance requirements. They over-configure to be safe: you get more grief for a slow app than you do for spending too much.

These issues are due to the political and cultural context that has developed around application deployment in the enterprise. Moving from disk to flash isn’t a fix.

The least convincing argument in favor of the all-flash data center is that the combination of the duplication and/or compression will shrink the data sufficiently to be competitive with the cost disk. The bottom line is that any technology that can reduce the cost for flash can also reduce it for disk.

Nimble Storage and Tintri – among others – already do this. Their hybrid disk/flash arrays features always-on data compression. It works at wire speed and most users probably have no idea that it even exists.

The StorageMojo take
While someone, somewhere, will undoubtedly invest in an all-flash data center, very few businesses will go in that direction in the next 10 years. The industry is responding to the impact of flash and public cloud storage by rearchitecting storage to incorporate flash where it makes sense, and by radically reducing the cost of high-capacity, high-performance storage that incorporates disks.

The cultural issues that make data center storage expensive aren’t easy but they are fixable. New storage systems that stress commodity hardware and scale out architectures can be looked at as horizontal layers rather than vertical stovepipes.

And the simple truth is that most data is only rarely looked at. Efficient erasure codes and geographical data protection means that backup is more likely to disappear than disks.

Courteous comments welcome, of course. Not that I’m suggesting that backup will disappear any time soon: as long as digital media are flaky we’ll need it.

{ 12 comments… read them below or add one }

gchapman June 25, 2013 at 7:10 am

The other aspect of dedupe/compression is IOPS savings, especially within the dedupe realm, if I dont have to re-write data then my requirements for write performance goes down considerably, augmented with flash or ample cache read performance becomes a trivial matter.

As more and more storage is fronted by virtual environments, inline dedupe/compression will become a bigger factor, and post process actions will become insufficient.

Disk isn’t dead, its just the bridge till the cheap high performance, high capacity unicorn is found.

Lee Johns June 25, 2013 at 7:39 am

Good points here and I agree with them all.

If you want to understand the realities of what it would take for SSD to replace HDD you might want to read this article by Tom Major in Sys-Con “SSD vs. HDD. The battle for the future of Storage”.

Tom was a Vice President at Seagate and is now President and COO of Starboard Storage Systems.

Starboard Storage (Who I also work for) also delivers a hybrid storage solution with inline compression on the SSDs and with that architecture we have demonstrated that we can get as much as 400,000 IOPS from a system costing less than $50K. Blog & video here

Now these are IOMETER numbers and not a real world environment but it shows the power of the architecture which has an innovative HDD storage pool that does not create the stranded storage of older array architectures and is multiprotocol so you can consolidate workloads and therefore use less SSD resources that folks like Nimble and Tintri (Who you mention) who are more restrictive in the protocols and workloads they support.

As you say the all flash guys are prone to comparing to old array architectures with no advanced data management technology. The reality for customers is that there are really great solutions that could dramatically reduce there costs and improve their performance out there and they are using a combination of media. HDD is not going away anytime fast. The Macroeconomics don’t lie.

Fazal Majid June 25, 2013 at 8:00 am

It depends on your workload. I think for OLTP the tipping point is already here. At my company we’ve been replacing 24-drive 10k SAS disk shelves with a single Intel 910 PCIe SSD (the disk shelves have been reassigned to secondary backup/archive storage). Redundancy is handled through replication, not RAID. The cost is about the same, and performance tripled in terms of throughput.
The superior IOPS and latency of SSD mean that you get much better throughput on OLTP apps, which are often bound by database round-trip latency and lock contention.

TD June 25, 2013 at 8:25 am

From the YouTube website:

100 hours of video are uploaded to YouTube every minute

Can flash keep up with that capacity?

David Magda June 25, 2013 at 2:44 pm

To add another data point, current roadmaps (from WD and Seagate) have HAMR encoding for hard disks coming out in 2014.

Introduction densities would have four-platter HAMR drives be 6.4 TB in 2014, with theoretical limits of the technology have 60 TB 3.5-inch drives ten years out. I’m dealing with a lot of DNA/genetic bits, and bulk storage isn’t going away for us any time soon. Of course everyone wants things to be fast as well. And don’t forget cheap….

While we’re talking about “hybrid storage pools” (which the ZFS guys were kind of prescient about), we’ll continue to see the growth of “hybrid drives” as well.

Hopefully though, instead of having the flash hidden away and only seen by the drive firmware, it will be exposed so that the OS can manage it. 1-8 GB of SLC, 32+ GB of MLC, and lots of gigabytes on platters would probably be a really good combination for tiered storage what would make things fly. Put all of that behind a standard SATA/SAS port and a port multiplier and let the file system folks go wild.

I think this last point is key, and would be especially useful to the enterprise-y folks to sell to the large storage vendors. I think whichever disk maker makes a transparent platter/SSD combination is going to make a lot of folks very happy.

Brad Shaw Borge June 25, 2013 at 3:38 pm

It’s true ssd’s are lot more faster. Most gamers today would prefer ssd than hdd. I would suggest of this site also lot of info too.

JR June 27, 2013 at 10:11 pm

I see a lot of similarities to the end of the CRT.

LCD monitors came to market at a huge price premium to CRT’s. They offered lower energy use and much smaller sizes. At the time the pundits said “They will never display CRT’s as the price delta is too large and the CRT factories are bought and paid for and can always underprice LCD’s.”

The conversion of the 99.9% market from CRT to LCD’s in the computer monitor space happened in under 5 years. The commercial mainstream use of CRT’s also transitioned to LCD/Plasma/etc in a couple years more. But today CRT is gone except very specialized high end applications – and that is rapidly dying as well.

Flash will do the same. It is only a matter of time until spinning rust is a distant memory. Apple is already pushing this on their laptops – so the high end consumer space has already jumped.

Another factor is improved reliability – as a non-mechanical system flash should be more reliable. So in cloud-scale applications where 3x replicas (or more) and advanced error correction codes are required to get the reliability with flash it might be possible to need less. This will just help justify the higher price of flash which will ultimately accelerate it’s price decrease.

The clock is ticking and spinning rust is on borrowed time.

Arun Balakrishnan June 28, 2013 at 12:32 pm

I work at Symantec and we agree that the $/GB for HDD is still much lower than that of Flash. If you assume 10-15% of data in a datacenter is “Hot”, to maximize the ROI on Flash, one needs to target just that hot data. If you are buying Flash purely on $/IOPS, then warm or cold data storage on flash will not meet the ROI metrics. Here’s where you can look at using disk arrays, JBOD, tape, or public cloud, depending on business needs.

That being said, one also has to consider that hot data is transient, meaning it may be hot for a week, cold for month, and then hot again at the end of the quarter for billing and business intelligence. To run massive BI reports, Flash can drastically reduce the analysis times, but do you want to keep all the data on Flash just in case? This is where seamless data movement, storage tiering, and caching can help maximize both Flash and HDDs to drive down costs while realizing the performance gains Flash affords.

So, while HDD still has a definite place in the enterprise data center, with intelligent software to ensure the right data is on the right medium at the right time, one can build highly predictable, highly performing system without compromising on effectiveness of storing longer term data.

DH June 28, 2013 at 4:26 pm

We own an SSD array that replaced a traditional array from a leading vendor. Both are good products but the SSD in our situation was put in place to handle the workload of a large VDI rollout in addition to housing all of our virtualized servers as well as file shares.
24U to 6U
Performance is great
Set it and forget it – this may be a function of design choices rather than SSD technology but welcome none the less

The kicker is the price quote we received for a new array from the same leading vendor came in very close to the cost of the SSD array

JP July 11, 2013 at 7:24 am

The Starboard numbers may be great from an IO perspective, but it will be an uphill battle for Starboard to succeed after letting it’s sales team go and putting it’s self up for sale. Especially when you have companies like Tegile, Tintri, and Nimble flourishing.

Chris Emura July 16, 2013 at 11:52 pm

“The bottom line is that any technology that can reduce the cost for flash can also reduce it for disk.”

I’d probably replace “any technology” with “most technology” though I’d agree that “any” makes intuitive sense at first glance.

Where this may break down spectacularly is in cases where the deduplication (or related technology) algorithms and data structures no longer need to be sensitive to the HDD seek penalty. An oversimplified example would be certain index lookups which would be way too expensive if done via random reads, but for which is no problem on SSDs.

Garrett Kajmowicz July 25, 2013 at 12:40 pm

NetApp Worker Bee here – all comments are my own.

There are a few other benefits of flash not covered here. Vibration resistance is one. Not a big deal for a typical datacenter. Those on ships or used by the military might prefer that as an option. Of course, this is why the iPod no longer uses mini hard-drives, too.

Another is noise levels. I’m slowly migrating my personal computing infrastructure to flash simply because less whirring is preferrable.

Something to consider is that all-flash isn’t the only way to deploy flash, either. One of our products is a card loaded with flash which is used internally to the filer as a persistant read cache, sort of like adding terrabytes of extra RAM to your filer at a fraction of the cost. Another uses a small portion of flash drives in concert with spinning media to create hybrid aggregates (probably similar to those mentioned previously).

Finally, you can do to the all-flash array in a host of different ways which have been covered here previously.

Leave a Comment

{ 2 trackbacks }

Previous post:

Next post: