PURE STORAGE: Are they the last man standing ? VFD3

During Virtualization Field Day 3 we visited the folks at Pure Storage in Mountain View.


Pure Storage are the guys with Orange colour-themed branding. See the gorgeous Lego-model Pure array on the table in the picture above just in front of Stephen Foskett.

They are also the guys who have received funding of $240 Million and borrowed Vaughn Stewart from Netapp. Somebody must think they’re pretty hot.

The Pure Vision

Pure Storage has a simple vision: to build something unique.

The key foundations of their strategy are to build an all-flash array that is high-performance, affordable, high-capacity, and back it up with outstanding support.

Before I visited Pure I didn’t really “get” the all-flash approach. I certainly would not have been convinced by it and this goes back to considering Texas Memory Systems 5/6 years ago. High-performance but uber-expensive solutions. Point solutions for problem applications, but out of reach for most.

Balancing performance and cost as usual

When I went to Pure I wondered how can they balance capacity, performance and cost, and what about wear levelling and Flash wearing out. If you don’t use Single Level Cell (SLC) Flash, will the quality be good enough to support the Enterprise. Many Questions.

And will they do it so well that EMC or another Big Iron guy won’t eat their lunch through sheer market dominance. My expectation was like a visit to a Porsche factory. Great for Top Gear geeks but not practical for the rest of us.

Humble Pie 

Sometimes it’s nice to be wrong.

Having visited Pure I was impressed by their approach.

!! They have made a commitment to Flash storage.

!! No tiering

!! All Flash.

!! Optimisation using Dedupe and Compression.

My favourite aspect is their pragmatic approach to the use of Flash, and the “lack” of tiering. I’m not sure about tiering. My misgivings relate to block size and frequency of promotion of blocks from lower to higher tiers. It’s like “This IS a panacea” – a perfect cure for all. Hmmm

Tiering is one of those solutions that involves a rule-of-thumb. Here’s an example:


10% SSD, 20% SAS, 70% SATA.

Not very scientific is it ?

My biggest concern is the block size. EMC FAST can use a 1GB “slice”, HDS used to use a 42MB page size on the VSP. How effective can these page sizes be ?. Despite the use of Flash we still write to array cache and it’s the speed of de-stage from cache to the fastest tier that benefits from Tier 0. So thinking tiering benefits front-end IOPS is a kind of ridiculous notion really.

Look at Pure’s data de-duplication algorithms which use 512-byte granularity. That’s a slightly different context but is higher overhead but ultimately better in terms of capacity optimisation.


Pure’s Flash Strategy

The starting point I suppose for Pure is the type of Flash we are talking about. In the available types on the market today, initial systems that were available from companies such as Texas Memory Systems were based on Single Cell NAND flash technology.

SLC is best in terms of read/write efficiency but the analogy that was provided was that this type of media is like CD-RW drives, to perform re-writes to it you have to it the whole thing and wipe again.

In terms of the ability to store data, MLC, eMLC and TLC are denser multi-level cells and allow more bits per cell to be stored. This means more cells and higher capacity. MLC are not as efficient and lead to writing to a tighter noise margin. Pure explained that MLC is actually higher quality than they need.

They made a point that really stuck with me. MLC flash is used widely in devices such as phones that we expect to work for 1-2-3 years. So their strategy is super-smart!. Rely on the reliability of products made in high unit-volumes and leverage the improvements in manufacturing quality and reliability that are inevitable.

A typical introduction

Pure did admit there are only 10s’ of customers that run all-flash datacenters.

Most customers have some kind of hybrid model where the requirement starts off for an individual project or difficult application. This is where Pure will be a perfect first use case. In this case it is likely that the application has a very specific business requirement for better performance. The constrained system could be preventing the rollout of a new business application.

This is where Pure’s mantra kicks in.. Flat line is good …… for example better than 1ms


Now for the technical features:

  • 100% MLC flash
  • inline dudupe and compression
  • reliability … 99.999% achieved across all systems deployed
  • All inclusive software pricing
  • End-to-End guarantee
  • All software developed in-house
  • No training or PS services required

Pure are championing a better approach to storage acquisition cycles. From a 3-year refresh moving to a 5-7 year lifetime. Based on being able to reprice a contract as capacity is added, to leverage the reducing cost of storage and by association, the support costs. That’s smart too.

Storage Musical Chairs 

Cisco acquires Whiptail. No big deal. Who knows Whiptail ?

Violin Memory systems shows limited use case ?

EMC XtremeIO has been slated by analysts.

The rest …. meh….

Pure are working away, growing market share and doing the right things. I did kind of disagree with Vaughn Stewart regarding SCM (Storage Class Memory) and even standard server memory configs of 4TB+. This is where primary storage will live pretty soon.

Screen Shot 2014-03-11 at 21.16.34

They are to my mind the most credible All-flash vendor in the market today. Just look at these figures … So they may just have found themselves to be the perfect acquisition target for the leading networking company on the planet maybe ?.

DISCLOSURE: Travel and expenses for Tech Field Day – Virtualization Field Day 3 were provided by the Tech Field Day organization. No compensation was received for attending the event. All content provided in my posts is of my own opinion based on independent research and information gathered during the sessions.


10 thoughts on “PURE STORAGE: Are they the last man standing ? VFD3

  1. First let me say that I work at EMC. And also that I have friends working at Pure. Now that is out of the way, let me see if we can address a couple of misperceptions is this post.

    You say tiering does not work. Or at least YOU feel uncomfortable with the concept. More than 50,000 EMC FAST users globally may strongly disagree with you… Many say they could not achieve their business objectives without FAST. – But then you also are misinformed. Fully Automated Storage Tiering no longer does 1GB slices. It now is a much more granular 256MB. What is more is that EMC’s midrange VNX – which is the product category Pure fits – also employs FAST Cache which operated a page size granularity of 64MB to cache data on FLASH.

    That’s the technical details. What is more exciting is to ask WHY? Because most workloads and especially hundreds of mixed workloads on a single array, exhibits skew. What is skew? – The short answer is that it is the relationship between IO activity and capacity. As it turns out, – based on of thousands of real world customer cases – roughly 5% of data is highly active at any point in time. 95% of data is not. By dynamically moving hot data to FLASH, the overall capacity cost can be kept below $2/GB at scale. That is before de-dupe. If one has data sets that lend itself to deduplication, then the cost can be reduces to below $0.50/GB. That is the power of tiering: dramatically lower cost! A FLASH centric strategy below $0.50/GB is impossible to do with ALL FLASH today. Even Pure Storage readily admits that.

    This is not to say that we at EMC are against all FLASH. We ship more FLASH that any other storage company on the planet! And EMC has led the industry in FLASH adoption since 2008. We also just launched XtremIO, our built for FLASH horizontally scaling multi-node AFA, which has been extremely well received. In addition EMC are working with thousands of customers that deploy critical workloads on ALL FLASH in our VNX and VMAX storage systems.

    Bottom line is that FLASH is the new disk and disk is the new tape. We get it! As the market share leader, EMC is leading to shift to all FLASH for ALL active data. Dormant data on the other hand is most cost effectively stored on high capacity mechanical disk drives.


    1. Hi Denis,
      Firstly I want to thank you for your comments. There aren’t enough people commenting on blogs these days.

      I believe I’m entitled to take a pop at Tiering. Just as it is for you to disagree, which I entirely respect by the way. And some of the figures I mentioned for 1GB slices were actually quoted by a customer architect who I really respect, which I suggest kind of proves my point.

      I would point out that tiering was a small point of the article which I will certain write about in the future. Particularly how it plays with vSphere SRDS and SIOC. My personal view is to prefer to empirically size the storage solution. FAST and equivalent solutions remove that control to some degree due to a lack of a “guarantee”. I’m impressed that EMC have 50,000 FAST customers but that doesn’t surprise me. It’s a well-oiled Sales machine with rock-solid software and engineering to back it up.

      In terms of my motivation for writing those couple of sentences, I know how to size a platform for 50,000 SQL IOPS. I know how you would SCALE that platform in increments of 1000 IOPS for Data and Log, without using Flash. Having been through the challenge, and seeing the response of some vendors, I think it’s useful being able to do that without resorting to using Flash.

      I also think my point about front-end IOPS and tiering is entirely correct. Unless I’m mistaken we still use cache for writes, and under non-contended circumstances, de-stage happens (normally) before we hit the watermark, before cache write-pending rates become a problem. In an array without tiering that is properly sized in terms of cache, latency can be sub-1ms. I’ve seen this myself for very large workloads where no Flash was used, over a 6-12 month sampling period. That’s because it was properly designed for performance, capacity and scalability.

      My feeling is that many performance problems that existed in storage arrays before tiering were down to bad design. That’s an opinion from 13 years since my first Clariion FC4700 install in 2001. Bad design doesn’t always lead to bad experiences, but that doesn’t mean it couldn’t have been done better.

      I also think bad solutions / architecture design cost money, which is an intangible part of the cost/GB argument and the inability to properly balance performance and capacity. And my other experience is that no customer I have ever spoken to has ever mentioned the term Cost/GB.

      My final comment is regarding my experience being at meetings with vendors pushing tiering where no performance problems existed. I’m not mentioning any names but I was not sitting at the table with EMC. I think if we can answer that question that would also be useful and would make this post even more worthwhile.

      All the best,

  2. Interesting article but more things to consider:

    1. IBM has the fastest flash array on the planet with no one even close in performance. They achieve less than 1/10th of a millisecond speed. This is because of the their purpose built architecture. They don’t use consumer grade MLC flash like everyone else; they use enterprise grade flash and even better yet they don’t use SSDs. They build their own Flash modules for pure speed and use parallel processing across all their 12 Flash modules with 4 processors and 2 FPGAs per module.
    2. Flash is first and foremost about speed, then RAS and they have it covered from both ends.
    3. And as far as enterprise data management, they use battle proven software on their integrated V840 that not only virtualizes their own flash array but can virtualize anyone else’s arrays even disk. EasyTier, Replication, Thin Provisioning, SnapShot, Compression are all available as well..
    4. In the first 6 months of buying Texas Memory, IBM has sold over 1,000 new customers and the list is more than impressive. The largest companies in every market sector are running their heaviest workloads on mission essential applications and are ecstatic about the results. Got to youtube and search for IBM FlashSystem and see the numerous customer testimonials. For Example, http://www.youtube.com/watch?v=vRBZaARUW9g
    5. IBM is investing $1B in Flash development, who else can even claim 1/10th of that. Almost all their competitors are startups with Venture Funding and some of which are laying sales people off already, like Pure Storage and Violin Memory. IBM is not only the top vendor, but also they are the least risky. They are over $100B dollar a year revenue company and is committed to providing the highest in performance, availability, reliability, serviceability, ease-of-use, scalability, investment protection, and manageability.
    6. Is their really any other choice?

    1. Hi Ron,
      Many thanks for your comments and for the information. I currently work with an IBM partner who has seen success with IBM flash storage. From what I know it was for very intensive workloads that had become problematic. The flash systems removed the bottleneck which concurs with your points.

      However, in terms of comparing …. I’m not sure whether comparing what was Texas Memory Systems (and their current incarnation) and Pure Storage is an apples with apples comparison ?. What’s your view on that ?.

      I don’t normally quote cost/GB and other Gartner/IDC metrics. However, in terms of my post, part of my initial reservation regarding Pure was due to the fact that the cost per GB for Flash would be off the wall to allow it be considered for widespread deployment across an enterprise. If flash has to be ruled out due to cost, then it’s likely going to be a point solution, part of an automated tier solution, or presented in conjunction with lower performance tiers (maybe via IBM v7000). And having previous looked at Texas Memory systems a few years backs, cost was the prohibitive factor, although that was 4-5 years ago. And I think that’s why it was never adopted in large volume. Also the capacity was not great considering the cost.

      The fact that it had all the bells and whistles and was using enterprise grade SLC instead of consumer grade MLC was what made it prohibitive. I don’t think anybody’s doubting the performance, but that’s only one element of solution requirements.

      So I can’t comment on current cost, and I don’t think Pure are trying to achieve 1/10 ms, although someone from there might correct me if I’m wrong. They are about balancing performance at about 1ms which is excellent in anyones book, with capacity (via deduplication and compression), so you can deploy it in the Enterprise. So that was kind of what surprised me regarding their preference for commodity grade MLC. And they state during the session that they have had 5 SSD failures in total since launch. They’re Pure numbers, not mine.

      There are other solutions out there now such as COHO that can also do 1/10 ms using PCIe based Flash too, but are designed from the ground up to be very modular and scalable (180,000 IOPS per system) as well another platforms.

      If you can furnish more detail regarding comparing Pure with IBM Flash I’d be very interested in further discussing.

      All the best,

  3. Hi Paul,
    It is an apples to apples comparison. If you look at the compression IBM provides as an option on the V840 (the new incarnation of the Texas Memory product), you will see the effective usable capacity to be as high as 5X. Considering this, the cost per GB goes down dramatically and competes with anyone in the market for Tier1 disk replacement.

    Also, remember Pure is using Samsung SSDs, the same ones in laptops and they fail quite often. With compression, With the added functionality that IBM added to the Texas Memory product, it is not just for heavy workloads anymore but for all data in the enterprise and they keep their high performance advantage even when using compression.

    A single 2U FlashSystem 840 providing a usable capacity of 40TB after RAID5 but before compression achieves 1.1M IOPs and all this at 1/10th of a ms of latency.

    I would urge you to reach out to IBM and write another article comparing the two.


  4. Morten,
    That is what you call PURE spin. Pure uses MLC which wears out after 3000 to 5000 program/erase cycles and eMLC which IBM uses lasts 30,000. You just can’t argue with the technology specs. Also, Pure just laid off 2 sales reps here in Atlanta and 1 in North Carolina for lack of revenue compared to IBM which just keeps adding sales and technical staff across the country because of their success. Pure’s use of consumer-grade components and inferior speed compared to IBM’s enterprise-grade components and superior speed is an issue with enterprise customers. Most large enterprises just don’t trust Pure Storage which is shown by their wallets. Just look at their client list compared to IBM’s. IT staff loses their job when an application is unavailable do to a storage outage and IBM opted to buy Texas Memory for a reason. I would urge anyone looking at shared flash arrays to get in contact with their IBM Flash Specialist to hear the whole story.

  5. @Ron

    I am not saying what IBM’s products can or can not do. I am only saying that you comment about Pure Storage’s SSD “fail quite often” is incorrect.
    Pure Storage have sold over 1000 systems (each system have minimum 22 SSD’s) so at minimum 22000 SSD’s have been shipped (the real number is properly more like 50000) and only 5 have failed so far. ONLY 5!

    I have for the last 3 years used Flash in a enterprise environment from a lot of different vendors.
    – Enterprise SSD’s both SLC and MLC
    – PCIe both SLC and MLC
    – All flash arrays
    Our 2 Pure Storage arrays have never had a failed component, all the others have at some point failed.

    Just to be clear I don’t think it is an issue that redundant components fail from time to time, as long as the contract covers it and the failures don’t impact performance.

  6. Wonderful blog! Do you have any suggestions for aspiring writers?
    I’m planning to start my own website soon but I’m a little lost
    on everything. Would you recommend starting with a free platform like WordPress or go for
    a paid option? There are so many options out there that I’m completely overwhelmed ..

    Any suggestions? Thank you!

Leave a Reply

Your email address will not be published.