Hitachi Hard Drive, picture taken at Hitachi uValue in 2010

This week we’ve seen the announcement of 14TB 3.5” hard drives from one vendor and 12TB from another. It’s hard to imagine the concept of 14TB in a single hard drive when ten years ago we were looking at only 1TB devices. However, these devices have a problem. The throughput and latency has remained relatively constant during the last 10 years. 7.2K RPM drives from ten years ago had a throughput of around 120MB/s at best (sequential). Today, we see around 220-230MB/s, but individual I/O latency for the newest drives (which use SMR) is about 50% worse for writes than they used to be.  It’s time hard drives were permanently relegated to the archive tier with tape.

The Performance Problem

Hard drives have been an amazing technology.  From the origins at IBM in the mid-1950’s, drive manufacturers seem to have been able to overcome seemingly insurmountable physical challenges.  As I discussed in a recent post, the industry hasn’t quite reached 100TB drives, but we’re moving there quickly.  However, we continue to scale capacity exponentially without the same increase in performance.

The market for hard drives has split in two.  Capacity drives continue to push on with capacity increases, albeit with worse latency.  At the high end, performance drives have effectively stalled in capacity because the market is moving to flash where SSD capacities far exceed hard drive models.  The only possible reason for retaining hard drives (in either category) is cost.

Cloud Storage

The divisions in the market are highlighted in the recent podcast we recorded at Storage Unpacked.  We talked with Andy Hardy, Regional Vice President at Nasuni.  I haven’t spoken with the company in some time and in fact haven’t written about them in perhaps 3-4 years.  I’ve posted some links at the end of this post, but as background, Nasuni is essentially a global NAS solution, deployed on-premises with the data stored in the public cloud.  Andy raised a number of interesting points during our podcast.

  • Dollars are with the IOPS – this is obvious from a tiering perspective, but even more accentuated with the deployment of flash.  I/O performance costs money, so optimise the data on fast media.
  • Businesses have a 80/20 rule for data – 80% is unstructured (object/file), 20% structured (block).  It’s likely this ratio will change in favour of more unstructured content.
  • The CFO doesn’t like unnecessary spending.  For global companies, ensuring IT is deployed in the right place can be hard and wastage exists.

Have a listen (and subscribe) to the podcast for the full story, but we can see Nasuni as ratifying the “Flash and Trash” message that Enrico Signoretti highlighted a few years ago.  Of course this isn’t suggesting that data on spnning media is trash, but has a different way to be treated.

The Architect’s View

Nasuni isn’t the only cloud storage vendor in town, but their experiences show that it is possible to get data out of the data centre and into Public Cloud.  Even with on-premises object storage, most data can be served out of cache, with HDDs providing no more than an active archive.  I wonder how much longer the HDD manufacturers will keep selling performance devices and how much of that business flash continues to take.  If I was doing business transformation today, I would be moving away from on-premises deployments as quickly as possible.  The only exception would be where data needs to be onsite for compliance.

One last thought; next month I will be at Storage Field Day 14 listening to INFINIDAT, a company that has built a modern storage array on hard drives.  InfiniBox bucks the trend the industry is taking for performance IOPS platforms.  How will they continue to grow as hard drives slow down?

Further Reading

Comments are always welcome; please read our Comments Policy.  If you have any related links of interest, please feel free to add them as a comment for consideration.  

Copyright (c) 2009-2017 – Post #3B6E – Chris M Evans, first published on https://blog.architecting.it, do not reproduce without permission.

We share because we care!Share on Facebook4Share on Google+0Tweet about this on TwitterShare on LinkedIn98Buffer this pageEmail this to someoneShare on Reddit5Share on StumbleUpon0

Written by Chris Evans

  • John Martin

    Hey Chris, for the most part I agree with you … but there are still a few places where spinning rust still makes sense for production workloads which are biased towards read-mostly sequential throughput, such as video streaming, batch oriented analytics etc. For those $/MB/sec still favour the HDD. The SPC-2 benchmark results are a pretty good indication of that. I also suspect that the percentage of unstructured data will rise to around 90 – 95% as the majority of new data over time is likely to come from sensors etc .. though if you put that into something like spunk or elasiticsearch does that still count as unstructured in the sense that you were using ?

    • John, fair point, do you not think that’s just a cost comparison though? Given the ability to have flash media costing the same as HDD, you’d choose flash, right?

      It’s an interesting question on Splunk (I assume you meant) or ElasticSearch as to whether this is unstructured and perhaps it’s a grey area – semi structured?

      In lots of ways I was indicating that in the data centre we could eliminate a huge percentage of data to archive or cloud, and (aside from collection) HDDs could be usefully replaced. It’s a debate to be sure, one that we see moving inexorably to flash.

      BTW, I didn’t want to say that HDDs were “dead”, like we hear so much about tape, just that we’ve had “peak HDD” if that could be used as a term!

      • John Martin

        If you could get an SSD down to the same $/GB price as an HDD then it would be a no-brainer .. better density, thermals, power, reliability, performance. BUT, based on the projections I’ve seen (which are as good as one is likely to get), the price differential between SSD and disk is likely to remain about 8-10X for the next decade. Dedup and compression and operating costs can bring that gap down a LOT, but there are workloads where that isn’t true, yet.

        I’ve even seen projections where the $/GB of “archive” disks gets lower than tape, part of me remains dubious about that, but if you throw in operating costs then I think that’s fair.

        As far as “Peak Magnetic Disk”, I completely agree, that time has already passed, the 15K drive make no commercial sense, the 10K still has some life in it, but not for long, the 7.2K SATA disk may live on for a while, but in the end I suspect the only disk will be running at 5KRPM or slower and will spend much of its life in a spun down state.

        Jim Gray may have called it a little early, butI think he was the first one to say t”Tape is dead, Disk is Tape, Flash is Disk and Ram locality is kind” .. https://jimgray.azurewebsites.net/talks/Flash_is_Good.ppt

        a little over ten years on and the industry is now at a point to prove him right.

  • Well, HDDs are making their last stand in the capacity storage market, which is really object-based storage. You could have performance tiers in OBS where warm or active archive data is kept in one tier and another tier where cold or deep archive data is kept. The performance characteristics of the drives could also be different. Use PMR drives in the higher performing tier and SMR drives in the lower performing tier where it is mostly write once. There should also be a focus on energy consumption. Helium filled HDDs use fewer watts than air filled HDDs. Some OBS vendors have also figured out how to spin down a certain percentage of their HDDs until data is requested, which would be an additional energy savings.

    The rapid advance of capacity of flash means you should plan on deploying it in the warm and active archive tier and eliminate the use of HDDs. That said, the current NAND flash market has some production constraints that might take several years to overcome while new manufacturing capacity comes online but it is important to consider it.

    Be careful when choosing public cloud storage use cases because most organizations don’t have enough internet bandwidth to retrieve a lot of their data from the public cloud very quickly and additional fees will be charged if you do. Better to keep cold and frozen data in your data center. Building an object-based storage cluster is not rocket science.

    • Tim, good points around the environmentals of the drives. As you say, helium filled drives have lower power consumption as an example. Now, I wonder in the SDS world (especially vendors with object platforms) whether there’s enough knowledge about the technology features to be able to automate that kind of thing? We have multi-dimensional axes to consider.

      Bandwidth to public cloud is definitely an issue. And of course egress charges. You’re right about the ability to build out object stores. It’s easy to do (in fact I built a NooBaa object store this afternoon in about 30 minutes). The question is whether IT orgs want to be bothered with it any more!

      • John Martin

        Given that a lot of my time is spent promoting my employers object storage offering, I have some good insight into what kind of research is being done to leverage spin down, unfortunately almost all of it is NDA. The biggest challenge will be at the application layer where you need to make sure there is a tolerance for time to first byte which has to be more than the spin up time for an idled disk. Glacier and other deep archive offerings from the other hyperscalers may encourage application developers to embed service level objectives when create their objects. If that does happen then the policy engines built into the object system should be able to do the right kind of intelligent placement that makes this possible. Even today StorageGrid can place objects onto tape using its archive node feature, so using spin down wouldn’t be that much of a stretch.

        As far as implementing an object store, I agree it takes very little time to get the a basic S3 store up and running, but if you’re looking at something which you can scale to significant capacities that is exposed to the outside world and / or spans multiple datacenters (which you’d want for resilience), then getting everyone to agree on DNS, and firewall and load balancer settings can require a few days of meetings and design discussions, and then the application integration and tuning can also work out to a few days of work as well. It’s not rocket science, but a full production multi-petabyte capable object store spanning geographically dispersed datacenters does need some reasonably careful implementation planning.

        • Well, I was probably a little glib if we are talking about standing up an internet-facing object-based storage infrastructure complete with DNS and all the trimmings. If you have a reference architecture from the OBS software vendor and if the OBS software vendor incorporates all of the reporting, monitoring, management, and QoS features you are likely to need in a multi-tenant environment, then it will take some time, but not an unreasonable amount of time to put it together.

          • John Martin

            No, not an unreasonable amount of time at all. If you’ve got all the information to hand and have the blessings and co-operation of the network and security teams, then implementing a multi-petabyte object store can (and has) been done within a day, and then hooking up and testing applications, and testing failure scenarios (disconnecting cables, powering off racks, deliberately breaking routing) takes a few more days while everyone gets comfortable with what happens and signs off on test cases and service level impact statements. But for a trivial example you can get an object store running on your laptop in less than an hour.