An SSD speed jump by four-times? Not so fast

An SSD speed jump by four-times? Not so fast

Summary: A technical paper this month outlined a creative method for making SSDs four-times faster (sometimes). But the headlines ignore the paper's nuanced results and its creative method for making SSDs faster, more efficient and longer lived.

SHARE:
TOPICS: Storage, Software
17

As reported by Tech-On!, Professor Ken Takeuchi of Tokyo's Chuo University, presented at the 2014 IEEE International Memory Workshop in Taipei this month a paper (not available online) "NAND Flash Aware Data Management System for High-Speed SSDs by Garbage Collection Overhead Suppression."

That's quite a mouthful, so let's unpack it. The "NAND Flash Aware Data Management System" is simply a system focused on NAND flash, the kind in our smartphones and SSDs. The "Garbage Collection Overhead Suppression" is the heart of the matter because garbage collection is a difficult problem.

Garbage collection 101

Flash has a very slow erase and write cycle – in the range of dozens to hundreds of milliseconds – much slower than hard drives. As a result the SSD needs to keep a supply of empty blocks to enable timely data writes.

As data is updated individual pages within a block of flash are marked invalid. When there are enough invalid pages in a block the garbage collection process writes the valid data to another block along with new data, freeing the old block for a fresh write. The flash translation layer keeps track of what goes where.

Because flash has limited endurance, the balance between garbage collection and maintaining a supply of free blocks is a delicate one. Too aggressive means premature flash wear out; too conservative slows down SSD write performance.

The new concept

What the professor appears to have done is taken advantage of the fact that many flash parts are capable of writing from 1 to 3 pages of new data inside an already written block. The software developed by Takeuchi's team, writes new data on pages located in a block due for garbage collection. The newly written pages are then rewritten to a fresh block, along with existing good data.

This enables parallel writes which are good for performance; reduces power consumption because block writes are power intensive; and increases flash life because garbage collection can be less aggressive and more efficient.

The professor simulated the impact of this technique on SSD write speeds and for all but two workloads the results were consistently in the range of 10 to 80 percent speed increase — not four-times. Here's the graph from the professor's paper, courtesy of Tech-On!:

Screen Shot 2014-05-26 at 7.32.14 PM
Simulated performance improvement.

The Storage Bits take

Around 10 percent to 80 percent speed increase in an SSD is a very good thing. But there are several caveats to keep in mind.

  • This is a simulation not a hardware prototype
  • The workload's performance varies depending on how much of the SSD capacity is in use
  • While the software could be part of a driver on the host system, it makes much more sense in the SSD

Bottom line: Don't expect to see this impressive speed up in the next three years. Flash vendors have other techniques for speeding flash performance.

Takeuchi is a brilliant young researcher whose work deserves more careful attention than outlets — other than Tech-On! — gave it. I look forward to seeing what else he develops.

Topics: Storage, Software

Kick off your day with ZDNet's daily email newsletter. It's the freshest tech news and opinion, served hot. Get it.

Talkback

17 comments
Log in or register to join the discussion
  • Although SSDs are plenty fast already.

    Although SSDs are plenty fast already - I get 10 second boot times with Windows 8. Very hard to do that with platter based drives. Faster would be great, but IMO we're at a pretty good place right now as far as speed goes. Still need to work on that price/capacity thing, though.
    CobraA1
    • Writing Is Slow

      Booting an OS is overwhelmingly reading data, not writing data. SSDs are plenty fast at reading. By comparison writing is rather slow. The strategy discussed by this article is about improving write performance.
      robradina@...
    • In 1983, a 5MB HDD cost over $1,000, or $200 per MB.

      Today, a 1TB HDD costs under $100, or about 10 cents per GB (a 2 million-fold in reduction in price over thirty years).

      SDD is rapidly approaching 50 cents per GB. The trend is in the right direction - and has reached a point that devices with SSD boot drives and solid-state storage in general has become common place.
      M Wagner
    • SSD drives, be careful

      As much as i would like to say the speeds are impressive, the only major drawback to a SSD drive is when it bricks, then the drive and data become useless. The problem with a SSD drive when it bricks is that there is no warning, so you have no way of knowing when the drive might die. At least on SATA hard drives you generally get a smart hard drive warning of impending failure to give you time to backup.

      I wish they would develop hard drive technology where say like 4 separate hard drive heads could read & write to the hard drive, that would boost it's speed by 4x.
      ajapierce
      • consumer SSD's brick, Enterprise SSD's are typically safe

        SSD's def. brick (consumer cloud backup is good solution for this!) but enterprise class SSD's have RAID like protections built into them. Each company has their own way of doing this, Fusion-io does it this way: http://www.fusionio.com/blog/adaptive-flashback
        bitlifter
        • onsumer cloud backup is good solution for this

          The Cloud man The Cloud. Really sick of hearing about the cloud.
          I have 6TB of data online at home. Data of all sorts. But a lot of multimedia files.
          I would rather buy a couple of more HD's and put them on the shelf than cloud anything. Cost, 200 maybe 300 bucks, monthly outlay, ZERO.
          ChicagoBob123
      • Raid1 is the answer but expensive

        If your drive was Raid 1 (basically two drives in mirror) then when one drives you will still have your data. The issue for SSD just like SCSI is cost per GB (GB's is the new MB anyway) . Currently SSD is about 5 times more expensive than a Hard drive if not more.
        4TB for < $200. If they come through with new tech to reduce the actual cost hard drives will be use for back up like tape was in the old days.
        ChicagoBob123
  • People who say SSDs are plenty fast already

    Remind me of the oft quoted (if wrongly attributed) "640K is plenty of RAM" statement by Bill Gates.

    But on another note, improving SSD speeds is futile when the SATA bus is already fully saturated by the most generic drives. More effort needs to be put into the next interface. I am not partisan in this. But I don't think SATA IV is a realistic option, same for SATA express. PCIe is the only credible bridging technology imo.

    I am not a fan of Samsung really, but they have somewhat clandestinely released a pcie ssd into the wild and this subversive move was very much needed.
    CornheadsBack
    • It seems to me that flash technology itself is in need of another look.

      I understand the concept that erase/writes are slower with flash than with HDD technology and that garbage collection techniques and improve that but what does that mean to the average user? The keyboard/mouse is ALWAYS the slowest "device" because the human mind has to think between keystrokes/mouse clicks. The next "slowest" device is the printer.

      The article suggests that SSD is dramatically slower than HDD - which would otherwise be the "next slowest" device!

      In any event, though, isn't the real issue MTBF (mean time between failures)?

      My impression is that the typical HDD has an MTBF of around 50,000 hours. For SSD, the MTBF is more like 150,000 hours.

      So, if running an HDD 24/7, we can expect a useful lifetime of around six years - not an unusual period of time that a user or a small business expect to operate a server. This gives us an idea of when we should plan on replacing them. By comparison, shouldn't we expect our SDD drives to last almost 18 years? (Far longer than any existing technology has been in place since the introduction of the PC itself!)

      Today's HDD technology costs about 10 cents per GB. SDD technology costs about 50 cents per GB and, based upon the estimated above, that comes to fives times the PER GB cost for three times the MTBF.

      I would conclude that this is a good time to started thinking about SDD for durable data storage. Right now, that means boot drives and drives holding data which is relatively static in nature. For data which is manipulated a lot though, HDD with reliable backup is still probably the more cost-effective solution.
      M Wagner
      • Many good questions

        I'll tackle a few.

        SSDs have a slow re-write cycle - slower than HDDs - but are very fast on random reads which, in most workloads, predominate. Therefore they are faster and they usually have higher bandwidth than HDDs, that get slower as they fill up.

        HDD MTBFs are in the 300,000 hour range. SSDs are probably a little more reliable, but they also fail, and in some different ways, such die failures and wear out.

        Since HDD capacity is so much cheaper, they will continue to rule where capacity, not performance, is key.

        Robin
        R Harris
      • re: It seems to me that flash technology itself is in need of another look.

        Reliability is always a tricky question with SSD's. A main factor that affects flash reliability is how often it's written to as well as the intensity as to which its written. It's very easy to burn out nand flash by writing too heavily to a single die and thus cause early die failure. However, if the drive spreads out the writes across many dies, then then life of the flash is extended. Each vendor has their own secret sauce for determining the best way to write to their nand. Also, if the SSD isn't overprovisioned properly and is constantly being filled up and erased, it will start failing much earlier than a drive that had the same amount of data written to it, but over a much longer period of time. The reality is that cost per gig is no where near that of traditional spindles (~10 cents/Gig) but cost per IOP (performance) is heavily favored towards the SSD market. How many sleds or even racks of DAS or SAN storage equals the performance of a single high perf. flash device? Personally, I wouldn't use SSD's for mass storage of data just yet. Spindles, while slow, are still able to maintain data longer without power, and cost per gig is highly in their favor. With 10TB spindles on their way, I like to think that spindles are the new tape format. ;)
        bitlifter
  • SSD. Always purchase a recognised name brand….

    Crucial, SanDisk and OCZ have proved to be excellent SSD Drives for myself.

    SSD is still an expensive upgrade but the benefits far outweigh the costs.
    5735guy
    • SSD benefits far outweigh the costs.

      "SSD is still an expensive upgrade but the benefits far outweigh the costs".

      Unless you are using an ASUS MB with Nvidia chipset.
      My ASUS P5N-D with a NVIDIA chipset cannot take advantage of the speed of a SSD because of incompatibilities.
      rocketman67
    • OCZ an "excellent" SSD ???? In what planet???

      Guess why the company is going under? OCZ make the worst drives you can get. Most of them breaking down within weeks of purchase and a very large amount of them be DOAs out of the box.

      Sorry, but the moment you claimed OCZ as a good brand you lost any credibility.
      wackoae
      • Re: OCZ....

        I can only say I have not had any problems.
        5735guy
      • OCZ is going under?

        Say what? Are you not aware that they were recently acquired by Toshiba? They are now on much more solid footing. Like many companies, they had some not-so-great products in the past, and also some very solid high-performing products. You are certainly entitled to your own preferences, just like anyone else. "Most of them breaking down within weeks of purchase and a very large amount of them by DOAs out of the box". That statement simply is not supported by the facts. You are living way in the past.
        renosablast
  • Thx for getting the real story right

    I always appreciate someone giving me facts as opposed to FUD ;)
    bitlifter