Home | Storage | Comparing and Contrasting All Flash Arrays – All Vendors (Updated)
Comparing and Contrasting All Flash Arrays – All Vendors (Updated)

Comparing and Contrasting All Flash Arrays – All Vendors (Updated)

11 Flares Twitter 3 Facebook 2 Google+ 6 StumbleUpon 0 Buffer 0 LinkedIn 0 11 Flares ×

Don’t you just hate it when someone beats you to the blog post you intended to write.  Following on From Vaughn Stewart‘s Pure blog (Comparing and Contrasting All Flash Arrays), Calvin Zito (HP, Around the Storage Block Blog) has updated the sheet and added the HP 3PAR 7450.  Obviously, it’s Calvin’s job to show HP products in a good light, so kudos to him for producing the updated version.  However I want to look at the bigger picture and show as many vendors as possible and expand on the figures being shown.  So here it is, my extended summary of all-flash arrays.

First of all, let’s qualify what’s meant by all-flash.  Version 1 of all flash arrays basically placed flash devices into existing hardware.  I’m personally qualifying these out of the data and only including products that have been specifically designed or modified for flash.  In the examples chosen, HP 3PAR 7450 was specifically modified (both hardware and software), as was the Hitachi HUS VM (FMD hardware and software changes).  Products like VMAX and VNX aren’t included as they don’t meet this category.

Next, I’ve extended the metrics being recorded.  I’ve added a section for performance (glaringly omitted from Vaughn’s sheet), plus some other categories, including those Calvin added.  I’ve also indicated whether the platform is a scale out architecture, which will be important going forward.

NOTE: The latest version of this sheet is now online at http://codex.architecting.it/index.php/Solid_State_Arrays_Features_Comparison and is updated semi-frequently.

As you can see, this is a work in progress and I have lots of gaps.  There are a number of issues here:

  • This data takes a long time to trawl for and interpret.  Vendors don’t use the same terminology.
  • Some references are textual in product descriptions and not explicitly placed into data sheets.
  • Some data is questionable.

What do I mean by questionable?  Well, there’s always a difference between read & write latency and between read & write IOPS, but some vendors don’t split out these two values.  Some figures, like IOPS, are based on one measurement; for example 4KB blocks. Others, like bandwidth, are based on 1MB blocks and don’t make it clear whether this is sequential or random throughput.  Not all vendors are showing RAID before/after in their figures (presumably because they are variable).  Nimbus Data’s latency figures seem remarkably good compared to others on the market.  NetApp don’t even provide latency figures and only say “sub-millisecond”, which has to be taken as meaning 1 millisecond.

What’s the point of this kind of comparison sheet if the data isn’t consistent?  Well, having something in one table makes the outliers easier to spot, which in turn allows customers to question vendors on how their figures are calculated.  Presenting as much data as possible means potential customers can match up their requirements with product offerings.  It’s all about the requirements.  At least with some baseline for comparison, a single table is a good thing.

The Architect’s View

This is a version 1 “work in progress” spreadsheet, taking data from existing public online sources.   I would welcome any vendor who wants to offer corrections or additions, with the only caveat being that the data must be publicly available (please provide link), so anyone reading this blog can check the validity.  If you want to reference this chart, I provide no guarantees of the accuracy, but you may reproduce it as long as you reference the source.  I’ve noticed a few sites reproducing some of my graphics without permission.

Related Links

 

Comments are always welcome; please indicate if you work for a vendor as it’s only fair.  If you have any related links of interest, please feel free to add them as a comment for consideration.

Subscribe to the newsletter! – simply follow this link and enter your basic details (email addresses not shared with any other site).

Copyright (c) 2013 – Brookend Ltd, first published on http://architecting.it, do not reproduce without permission.

About Chris M Evans

Chris M Evans has worked in the technology industry since 1987, starting as a systems programmer on the IBM mainframe platform, while retaining an interest in storage. After working abroad, he co-founded an Internet-based music distribution company during the .com era, returning to consultancy in the new millennium. In 2009 Chris co-founded Langton Blue Ltd (www.langtonblue.com), a boutique consultancy firm focused on delivering business benefit through efficient technology deployments. Chris writes a popular blog at http://blog.architecting.it, attends many conferences and invitation-only events and can be found providing regular industry contributions through Twitter (@chrismevans) and other social media outlets.
  • Kevin Stay

    Thanks for putting this together! It is even more clear why EMC hates Pure so much… I have no use for a ‘pure flash’ system, but after looking over this side-by-side matrix it confirms my feeling that (at least at this point) 3PAR and HDS are really the only serious considerations for enterprise/critical work. Particularly if you consider that neither is a flash-only one trick pony like some of the others here.

    • Plain old Greg

      I can only speak for the HUS-VM here w/regard to the ‘enterprise/critical’ use case. I whole heartily agree that HDS is head/shoulders above the rest. However, it will cost you a mint. Even after discounts an all flash HUS-VM @ 34TB useable FMD will run you north of $500K. That works out to nearly 15K/TB. Also, don’t plan on any integration with vSphere like vCenter, vCOPS, etc because the functionality is horrible and pales in comparison to what EMC, Netapp, and heck even IBM offer up.

      • klstay

        That price seems about on par with everyone else to me. Even with ‘enterprise level’ discounts (we are a NetApp shop since ONTAP 6 and probably will be for the foreseeable future despite my preference for 3PAR or HDS) a DS2246 shelf full of 200GB SSDs will put you back $45k and you would need at least 10 of them to get that capacity. Plus the controllers and the software which is where they really make their money. 26U of rackspace and the wiring nightmare that goes with it for what is still at it’s core an ALUA box? No thanks. Give me a mini-vsp in 5U and 12 of the new 3.2TB modules in 2U please. For block loads with proven reliability 3PAR is the only real competition to the shared memory ASIC VSP design today.

  • David Nicholson

    EMC doesn’t “hate” Pure. LOL. We genuinely believe we have a better product. They disagree. Some of my best friends work at Pure. As far as the spreadsheet is concerned, I would suggest that a “test drive” is often what really makes the difference. Get behind the wheel of an XtremIO array and see what it can do for your environment. The opinions of 80s sports legends and 90s boy band members might persuade some, of course. If you look to Ronnie Lott for answers about IT infrastructure, you may need more help than either Pure or EMC can provide, however.

    • http://thestoragearchitect.com/ Chris M Evans

      David

      Certainly test driving makes sense. However using the car analogy, we all narrow our new car choices down using vendor brochures etc, before deciding on which ones are worthy of a test drive. Without comparisons, we’d be out test driving Nissan Cubes and Porsches.

      Chris

    • Storage Sloth

      That’s funny. Given Ronnie’s track record of investing in the valley. Also, taking advice from EMC on the latest/greatest product is like Blackberry telling us that they’re still innovating.

      EMC still faces the innovators dilemma. How will you create and push for a “better product” while preserving revenue streams from VMAX and VNX sales?

  • http://blogstu.wordpress.com stu

    Chris – some challenges with this type of chart:
    1. not all features are equal (deduce and scale-out vary greatly),
    2. price isn’t considered,
    3. management and operations of the machine can’t be boiled down to a couple of check boxes either.

    That being said, thanks for helping to moderate the vendor battle, always appreciate your opinion (and if you’re at HP Discover in Barcelona next week, the guys would be happy to have you on theCUBE).

    • http://thestoragearchitect.com/ Chris M Evans

      Stu, you’re right, no discussion here on price, however that’s a tricky one. I bet none of the vendors will give a fair price list and as you know, prices can be manipulated by adding/omitting config items (I’ve reviewed dozens of vendor proposals that supposedly matched a required config). This is an understandable commercial approach esp when bidding for large purchases.

      As far as features are concerned, that’s where detail has to come in and would be the logical next step for me, comparing how well those features are implemented, or at least highlighting the differences. A good example would be RAID/Data Protection approaches.

      This definitely is about giving (prospective) customers a baseline on where to start – it can never be a 100% comprehensive document.

      Chris

  • Tyler Britten

    [Disclosure- EMC Employee]

    Chris,

    While the content may not have been 100% accurate, I agree with the gist of Vaughn’s original idea.

    The idea is that Flash is different than disk in more ways than just “it’s faster and more expensive.” It also has a much more finite life, and a much more complex write/overwrite requirements and performance impacts.

    This is where true “AFAs” are defined- they include technologies to address those above mentioned issues. For example, XtremIO includes always-on inline deduplication not just for the space savings, but also for the reduction of the number of writes to the SSDs- improving SSD life and write performance. Same goes for the data protection on XtremIO- XDP. It designed for reducing the write amplification of traditional RAID (1.2x), while providing double-disk faliure protection, and doing it in a space-efficient manner (8% overhead).

    Now I’m not trying to turn this in the an XtremIO commercial, but those type of features are what makes an AFA.

    That’s what makes me exclude the 3Par 7450, NetApp EF and similar arrays. Yes, HP has streamlined some of their wide-striping to be more SSD-friendly, but is that really their AFA pitch? “We can do 4k writes now!”? Same goes for the EF (Engenio) array from NetApp. Those boxes in general have been good performers for years while IBM and Sun were OEMing them. My question is how is throwing a bunch of SSDs in either of these really that different than the VNX-F? They still use traditional RAID technologies, don’t include any tech to reduce write amplification, etc.
    The “we’ve streamlined our OS for SSDs” is a pretty specious argument as most vendors have been and are doing that. You can look at a number of code (and hardware) upgrades specifically in the VNX and VMAX to handle SSDs more effectively. Would I call them AFAs? No.

    Their target appears to be the well-established (and pretty niche) “if I can trim 5ms off this DB job we make $10M more a year” market. If you’re in a market for that type of solution, why wouldn’t you be looking at someone like Violin who eschewed the drive form factor?

    A good counter example of introducing SSDs to an existing architecture is HDS- They appear to have thought “how can we effectively use Flash while maintaining our current architecture?” The FMS/HAF is a fascinating engineering approach and shows some thoughtfulness to the problems at hand besides “Hey this flash thing is getting big. Let’s take one of our arrays and throw some SSDs in ‘em.”

    And to your question of inaccuracies:

    1) XtremIO (and I would assume pure) should get a “Y” for zero reclaim- with inline deduplication, the zeros would be instantly reclaimed upon writing :)

    2) XtremIO does support double disk failure (http://www.xtremio.com/data-protection-2)
    3) The max usable post-dedupe on the XtremIO is 250TB, not 75TB.

    • http://thestoragearchitect.com/ Chris M Evans

      Tyler

      So this is an interesting point. Do you include vendors who market a product as all-flash into a list or exclude them because they don’t appear to meet flash requirements? Is it better to include them and highlight how they can’t meet some of the flash optimisations?

      How about including warranty details? If the vendor is prepared to give a 5 year warranty, would that be sufficient, or is there still an issue here because there could be more failures (and so risk) with some vendors?

      We could perhaps have the same discussion with the “scale out” term. Is XtremIO scale out if it can’t tolerate a node failure?

      BTW, I’m not giving an opinion on the above – I’m genuinely interested in other people’s opinions.

      The interesting thing is how the market has evolved. All flash was a good marketing place for startups. Being able to talk about that unfixable workload that could only be solved with a dedicated appliance gave them an opening. However the market has matured. Point solutions for high-performance all-flash isn’t a growth market – otherwise Violin Memory would have been doing much better.

      Customers can get access to flash at increasingly lower cost and either retain their existing technology framework (HDS, 3PAR) or look at new solutions depending on their requirements. For many, the disruption of change may not be worth it.

      I’ll look at the two comments on inaccuracies and amend accordingly. I will be putting a more dynamic copy of the table up on a Wiki which will let me edit it more efficiently and show the references.

      Chris

      • Tyler Britten

        Chris,
        I don’t disagree on that part- that “all flash” isn’t some magic panacea that cures all storage woes and that hybrid ssd/hdd arrays are “dead.” Clearly there is a ton of value for customers in the hybrid array space and it is important that vendors make appropriate use of them in the most effective ways possible (tiering, caching, etc).

        I personally feel we’ll be in the hybrid model for quite some time- but it may be phase change memory as the performance tier with SSD or HDD or some other new exotic technology as a capacity tier.

  • Pingback: Overview of purpose-built VDI solutions to accelerate storage performance | UP2V

  • Greg Knieriemen

    Good table Chris – and it’s good that you note that this is just version 1 so it can be updated as more information becomes available.

    Some vendors might object to a simple feature list but, as you know, every implementation starts with a requirements doc which is largely based on a feature list. While a buying decision comes down to many other factors, the always start with a basic requirement document.

    Great start and I look forward to seeing how this evolves.

  • Vijay Swami

    [Disclosure: I work for a national VAR that has business relationships with multiple storage vendors]

    Chris,

    This comment is a follow up to the twitter discussion we had yesterday.

    The main premise of my argument is that the checkbox feature list comparisons do more harm than good for -customers-. Stu alluded to some of the challenges in his comments, but to expand on them, I’ll give you a couple of examples of things I have run into in the field which might shed some light on the context from which I was speaking:

    1- I had a customer who switched from vendorX to vendorY storage platform. One of their requirements was snapshots. As it turned out, during implementation they could not leverage snapshots in the same way they were doing previously, and their entire business process for DB copy management between dev/test/production stopped working. Because all snapshot implementations are created equal.

    2- Deduplication of primary data. I had another customer who decided that they wanted deduplication of their primary dataset. They saw on a feature checkbox that it existed, sized & bought the array on that premise. As it turns out the deduplication on this platform had tons of caveats and wasn’t appropriate for the workload that was being targeted, and caused massive performance issues upon implementation. Because all data reduction implementations are not created equal.

    Now these are just a couple of examples, but they both are a result of the customer not doing the due diligence because they believe “all features by a name are created equal”, which certainly is not the case. It’s the same argument for API compatibility between clouds — just because the APIs between two clouds are compatible doesn’t mean the -implementation- is that same and you’ll get the same -behavior-. And that is really the crux of my argument around the usefulness of these things to customers. They don’t encourage them to ask the right questions to see if they are the right fit for their requirements.

    Secondly, there are things that show up on these checklists that I feel should be 100% abstracted from the customer. For example I don’t think that the type of SSD used should be of any concern to the customer as long as the AFA vendor is doing SSD management appropriately, and the AFA delivers the right consistent performance & capacity at the right price.

    That being said, I think you have done great work here to dig into all of the various AFA vendors and categorize them in your chart. I do absolutely think there is value in what you have done for the industry as a whole to provide a starting point for comparison.

    I suspect our opinions differ on the usefulness of these things to customers based on the different roles we play in the industry as well as my customer experiences. As a VAR, or trusted advisor to my customers, I try to arm them with the right questions and information to seek in order to make a decision, and I’ve just found that in my personal experience these feature checkboxes do more harm than good no matter who leaves them behind. In an ideal world, all customers would ask the questions beyond just a “Y/N” on a chart, but that isn’t what happens in the field.

    Hope this clears up my view point. The AFA market is very much in its infancy and I think we are all going to learn plenty about the market in 2014+.

    Thanks.

    /vijay

  • Pingback: Flash Point | Yagnavalky's Blog

  • Pingback: The Maturing of Flash Storage | Architecting IT

  • Pingback: The All Flash Array - Oh, What's In a Name

  • Pingback: The Maturing of Flash Storage

  • flashdba

    A couple of things: [Disclosure - Violin Memory employee]

    Firstly, the subject of Non Disruptive Software Upgrades. I do not believe that any of the flash vendors who use SSDs (as opposed to ground-up flash designs like Violin and IBM TMS) have the ability to upgrade the SSD firmware non-disruptively. It is easy to make the claim that upgrades are non-disruptive, but the scope is usually limited to just the controllers. Consequently I think it is a mistake to place a Y against the SSD-systems for this line. Many of these systems are critically reliant on the SSD firmware provided by the third party SSD supplier, e.g. XtremIO’s comments that they don’t do their own garbage collection. I can confirm that the Violin array can non-disruptively upgrade the firmware on the VIMMS (i.e. the flash modules) as well as the controllers.

    Secondly, you use three line items for capacity: Raw, Usable (Pre-Dedupe) and Usable (Post-Dedupe). For me it makes little sense to see a larger value in the Usable (Pre-Dedupe) column than in Raw. I think it would be better to change this to Pre/Post Data Reduction on the basis that the Pre column shows *guaranteed* usable capacity, while the Post column shows what may be achieved with the right kind of data.

    Oh one more minor thing. I think it would make sense to reverse the wording of the line item UPS Batteries because most people reading this chart will assume that N is a bad thing and Y is a good thing. In the case of needing a UPS it’s clearly the other way around…

  • Roger Hinson

    Can you post the “spreadsheet” version of this? I’m going through something similar and would like to have a good start.

11 Flares Twitter 3 Facebook 2 Google+ 6 StumbleUpon 0 Buffer 0 LinkedIn 0 11 Flares ×