Chad's view isn't and never will be independent while he works for EMC. His focus has to be EMC-centric and why shouldn't it be; after all they pay his salary every month. People need to read his blog posts (and everyone else's) in the context of why/how they are written and therefore take statements with a pinch of salt.
An HP exec has taken exception to an EMC president’s blog about all-flash arrays. EMC’s Chad Sakac blogged that: AFAs are not right (at least not yet) for workloads that require "classic enterprise array" data services like huge at scale replication (think tens of thousands of objects replicated, extreme replication …
Friday 6th February 2015 12:24 GMT Michael Duke
The same can be said for Johnson. Chads point is that a true AFA is not the same as a 3Par 7450 or XP7 with pure flash. There is a measurable latency difference between most of the current startup/ex-startup AFA's and a hybrid array that is not hybrid.
Even HP's tools model approx 1.5MS of latency in an AFA 7450 with the MLC / cMLC disks.
Compare that to 500us for Xtreme-IO / Violin etc. A 3x improvement in latency is significant.
Friday 6th February 2015 14:29 GMT burjoes
if they were down to 1 nanosecond and 50 nanoseconds, a 5000% increase would sound amazing - and of course, nanosecond latency would be! But an application actually benefiting from 1ms less latency at the ~1ms level is a different question.
While there may be some applications and instances where it is helpful, do you really feel that it is significant in a reasonable percentage of real world use cases?
Friday 6th February 2015 16:43 GMT philipclark
Reply to: burjoes
If that were true and storage was 1 nanoseconds to 50 nanoseconds, then the CPU or network would be the biggest bottleneck.
But that is not the case: storage is 1 millisecond to 10s of milliseconds, 10^6 times slower than nanoseconds.
That is why today, storage latency has such disproportional leverage over the application response and end-user experience.
[Disclosure: IBM FlashSystem employee]
Friday 6th February 2015 15:54 GMT Nate Amsden
My 4-node 7450 (while far from heavily loaded) is averaging around 0.4ms. The official HP tools say this configuration is rated for 100,000 IOPS @ 90% write with sub 1ms latency w/RAID 1 (my workload is 90% write). With dedupe on I don't know what the number is(I'm sure it's less), but my actual workload these days is in the range of about 6,000 IOPS(migrating from 3PAR F200 with 80x15k disks 100% uptime since it was installed just over 3 years ago), so it's good enough for me for a long time to come.
I chose RAID 1 because I have 4 disk shelves (including two which house the controllers), RAID 1 gives shelf level availability so I can lose a shelf of disks and stay online(not that it has ever happened to me). That and 90% write is hard on the backend for RAID 5. That and given the large SSDs capacity was not an issue.
The cost was good enough to easily justify this route vs a hybrid or all disk setup -- though I wish I could of had a 7440 instead of 7450(identical hardware) so I would have the option of running disks for bulk storage if I wanted. Perhaps HP will unlock this self imposed marketing limitation in the 7450 in the future I don't know..
Main reasons I went with 7450 vs others outside of 9 years of 3PAR experience is I wanted a true 4-controller system(mirror cache writes in event of controller issue which is important in my 90% write workload), I wanted a mature platform for running this $220M+/year of transactions.
EMC cold called me earlier in 2014 and I talked to them for a couple hrs, but I was never about to consider XtremIO it is too new(also too inefficient power/rack/etc wise at the time this environment had 2 cabinets now it has 4 which will last 2-3 more years - 7450 takes up 8U and I can grow to nearly 200TB raw flash without more space). I do like the prospect of being able to have upwards of 500TB of raw flash in my system if it came to that(currently have ~30TB, maybe go to ~45TB some time this year - don't really see going beyond 75TB in the next 3-4 years but who knows).
I've learned a lot about storage over the years, by no means claim or even try to be an expert, but one thing I've learned is be more conservative. So for my small environment I opt for a mature architecture.
Friday 6th February 2015 20:44 GMT Anonymous Coward
"Even HP's tools model approx 1.5MS of latency in an AFA 7450 with the MLC / cMLC disks."
Not if you know how to use the modeling tools correctly and even then the numbers are typically conservative. That's especially true at these latencies where there's very little room for error so YMMV on any platform,
Friday 6th February 2015 15:53 GMT virtualgeek
A little more dialog
Disclosure - EMCer here (Chad)
@ChrisEvans: Chris you are dead right. I absolutely disclose who I am, and my employer - always - exactly for that reason. Regardless what people may suspect, I'm not a mouthpiece, I always say what I believe, what I can defend - and periodically, I say things critical about EMC and EMC products.
**BUT** it's inevitable that based on my day to day work, my exposure to what EMC and our tech partners are doing - it will colour my view. I absolutely have a sample bias (I think everyone does). Everything I say should be viewed through that lens.
Now, on to the post itself (and I will comment on Chris Johnson's blog to make sure the dialog continues in both places):
I fundamentally put architectures that were designed in a period that pre-dates flash (true of VMAX, VNX, 3PAR, HDS USP/VSP/HUS) - all of which have designs (for better and worse) that presume a high-latency destage of IO (and therefore lots of cache, and other things) in a different architectural bucket than architectures designed presuming that every single IO will land on NAND for persistence.
YES, all of those "legacy" arrays have been tweaked and tuned for higher and higher NAND mix and support 100% flash configurations. VMAX3 for example was designed for almost 10x the IOPs mix of VMAX, and the caching algorithms needed huge re-writes for cache bypass and other behaviours when the IO is to NAND.
YES, all of those "legacy" arrays have many data services (like the ones I noted that DON'T exist in the "designed from scratch AFA" group).
NO, I don't think a hybrid, originally designed in the 1990s (and updated furiously since then) can be called an AFA simply with tweaks and 100% flash configurations. I think that's true of EMC. I think it's true of the rest of the industry (many of which have 100% flash configurations of hybrids - like HP).
Frankly, I think Chris might be making my point :-)
Friday 13th February 2015 04:51 GMT Anonymous Coward
Tuesday 10th March 2015 13:51 GMT Rulon
Re: A little more dialog
I work for HP Storage. Being a built-from-the-ground-up all-flash array by itself is not a feature or even an advantage. And I am puzzled as to why EMC people like to talk about it as if it were. Bottom line: if compared to an established architecture, XtremIO is slower, takes up more space, requires more SSDs, is missing key features, has a problematic upgrade path, and a possible HA hole in how it handles metadata, then what exactly is the advantage of being built-from-the-ground-up for flash.
Built-from-the-ground-up is only an advantage if it produces better results.
Friday 6th February 2015 16:04 GMT virtualgeek
Disclosure - EMCer here.
@NateAmsen: I'm sure I speak for HP (in a sense!) when I say: "I'm very glad you're happy with your choice!". This is why I think it's good that at the same time there are "net new" AFA designs, that the older, more mature Hybrids are getting tweaked for all-flash configurations.
For the EXACT reason you picked your 3PAR 7450 (maturity, familiarity, specific data services), we do all-flash VMAX3 and VNX2 configurations, and they are VERY, very popular, and have lots of happy customers - just like you are with your 3PAR hybrid, configured with all flash.
But - all sorts of things, including the IO path, and RAID configurations you describe (which would be similar to on a VMAX3/VNX2) are examples of things that, if a developer was assuming a 100% NAND persistence layer, they would do differently (and almost all AFAs do differently).
Not saying "100% flash configured hybrid/cached array = bad!". They have data services like the ones I mentioned and are the right answer for some customers.
I am saying that true AFAs (designed to NEVER have a magnetic persistence layer) are the fastest growing category in SPITE of none of them having VMAX3 or 3PAR like data services.
Friday 6th February 2015 16:14 GMT Nate Amsden
I have been interested in trying some of the newer things, I just don't have an environment where I can take that kind of risk. Past companies had a lot more gear to play with..everything is so efficient and important here though that I try to be careful.. We do have a Nimble array at another site (sourced by IT I wasn't involved in that).
But I suppose the advanced data services aren't all that critical - I mean I don't use a whole lot on 3PAR. I don't use replication, I don't use many of the more advanced things (same goes for most tech I use, for some reason I tend to stick to the core stuff which tends to be the most solid whether it be storage, networking, vmware etc..). But the maturity aspect was important obviously. I've had my share of issues on 3PAR over the years.. I didn't think they would make it into the all flash world, I've been amazed at what they have accomplished though.
I'm sure XtremIO can work for a lot of folks..same for Pure storage and others..
Friday 6th February 2015 20:41 GMT pd_p
Still using eMLCs?
HP 3PAR Employee
Is it the ground up design of XtremIO that forces it use eMLC, just like its VNX/VMAX brethren. If it hasn't figured out a friendly way to work with flash, I'd submit it is as legacy an architecture as the rest of EMC portfolio in some sense.
On performance, if XtremIO indeed delivers 500 micro-seconds (for what workload?), it would make it one of the slowest AFAs. By contrast, 3PAR 7450 can deliver as low as 200 micro-seconds for single threaded writes, and routinely delivers less than 500 micro-secs of latency for multi-tenant workloads.
Saturday 7th February 2015 19:33 GMT tom 99
Re: Still using eMLCs?
Happy potential customer of XtremIO here... it depends on price I guess :)
In Dec 2014 my company tested XtremIO minimal config (I believe it;s the smallest possible) -- 13 disks, with our application transactional workload, Oracle based. Although I can not say what the max storage performance is, because we saturated all server CPUs before killing the array, we managed to achieve 117k IOPS at about 0.4ms latency with 1) majority of 8kB block size, 2) read/write ratio about 70/30. According to array management software, we were at about 45-50% XtremIO utilization. With low I/O workload, the latency was floating around 100-200 usec, same block size. Someone has cited 6000 IOPS with 3PAR 7450. Sure, it's possible, but this is a workload a single SSD drive can do.
The performance is only half of the story. At the same time we had 7.2:1 data reduction rate (3 copies of our application and database, compression ratio was around 2.5:1, dedup ratio 2.8:1, some benefit from thin provisioning).
Saturday 7th February 2015 04:50 GMT Anonymous Coward
Saturday 7th February 2015 12:46 GMT Mark Burgess
For me the "pure" all-flash array is just not sustainable and I think the industry needs to get away from the thinking that there is something "magical" about an AFA - that is just marketing.
At the end of the day it is just hardware and software - I see no reason why a company who focuses on a small number of storage platforms cannot overtime optimise their platforms to work optimally with flash and HDDs - that way we avoid silos.
The problem for the vendors that have "pure" AFAs, that have no support for HDDs, is that they have no motivation to optimise their hybrid arrays as their strategy is all about claiming that a "pure" AFA is in some way special.
The traditional hybrid arrays will continue to get further optimised for flash over the next few years and we will reach a point whereby products like Violin, Pure, XtremIO, etc. will just not make sense.
I would just love for a "pure" AFA vendor to explain what their product does that a hybrid array like 3PAR, VNX and FAS cannot do (either now or overtime) - certainly we all agree there is a whole bunch of things a hybrid array can do that an AFA can not.
Will it be quicker for an AFA to have all the features of a established hybrid platform or the other way around?
Based on the current trajectory I would say the established hybrid platforms will catch up with the AFAs first and I do not think that will be far way.
Just my view of course and I might be proved completely wrong - only time will tell.
I wrote a detailed blog on this subject over at http://blog.snsltd.co.uk/does-the-all-flash-array-really-make-sense/ recently.
Saturday 7th February 2015 20:07 GMT Zerolab
What is EMC position on VNX-F that's is considered as an all flash by EMC, and as such reflected in the various IDC and Gartner reports?What would of been EMC position in the AFA MQ without VNX-F numbers and with disclosing that you shipped a product that could not perform SW upgrades without whipping data of the array? Maybe we are all missing the design to land on NAND persistence, or whatever that means....
Sunday 8th February 2015 12:09 GMT Anonymous Coward
The EMC guys keep trying to bring the positioning back to VNX & VMAX's ageing architectures vs 3PAR, that way they can avoid having to do a 1 to 1 comparison of 3PAR vs XtremIO both on paper and in the field.
EMC's attempts to contain this conversation by repeating these claims as often as possible should be pretty transparent, They are visibly behind the curve here in both the AFA and Hybrid space and are simply trying to skew the market to their strengths until they can claim come form of feature parity. At the end of the day all the whizz bang dedupe and compression features only actually benefit the Customer if they can reduce his purchase price vs the competition and that simple fact isn't being reflected in many of the AFA vendors pricing, especially given their lack of enterprise features.
The fact is that 3PAR has a completely different and superior architecture, both in software and hardware than either of the above legacy platforms (both 20+years). Neither does it have any specific dependencies of spinning disk, it's a clustered I/O engine that can function and function well with any type of persistent storage (SLC, eMLC, cMLC) both now and into the future, and crucially it also has field PROVEN data services.
Sunday 8th February 2015 14:48 GMT Anonymous Coward
Monday 9th February 2015 09:29 GMT klaxhu
Anybody smart enough will recognize it's time to change will throw their own products under the bus if they have another one, better one, to replace it. If you don't have that newer/better/whatever - of course you will find all reasons to hold your ground on the current platform. That is what is happening right now ... The AFA startups play big on "new product designed from ground up" and they are right to do so, carrying that legacy it's not as easy at it sounds when refreshing a product and you rely solely on new iterations of CPU's and platforms.
In the end, the flash disk in it's more purest form will win (think everything in dram type thing!), because there is no other way. The rest are just details ...Or do you still see today the vast majority of people using CD's or cassettes in their cars/homes? No ...you see streaming, yes ..not even mp3's anymore.
That is super-evolution!
We just have to wait and see, but in no version of the futures I see, there is spinning disk or any platform that was classically designed for it.
HP will change its story like they want once they acquire one of the AFA's out there. If they don't, they are just ridiculously stupid!
Monday 9th February 2015 10:25 GMT Anonymous Coward
How is that when HP offer AFA today that provides competitive price and performance vs every other AFA on the market...Why would they need to acquire to in effect take a step backward. EMC had to acquire because their existing platforms just weren't fit for purpose in an all flash world, if we' were talking rack scale flash then that's a different conversation and a much smaller market today, but for the majority of flash use cases HP have a very capable platform that isn't standing still.
Wednesday 11th February 2015 17:24 GMT Anonymous Coward
"However, EMC VMAX and VNX customers appear to be endorsing EMC’s point of view instead, possibly because it is easier for them to embrace XtremIO,"
I doubt it as 3PAR includes inbuilt "online import" which provides in array hardware based migration for both the VNX and VMAX platforms, all with minimal setup and online data migration. This makes it effectively easier to go EMC to 3PAR than it is EMC to EMC. I thinks it's more likely many EMC customers just haven't looked beyond the EMC bubble and with Xtremio EMC are still seeding the market with those try now buy later deals.
Thursday 12th February 2015 12:12 GMT Valdis A Filks
All of these factors were taken into account.
Not sure if you are a Gartner client, but if you can please read the Critical Capabilities for Solid State Arrays http://www.gartner.com/document/2834717, Also, please have a look at this post from Chris Mellor; http://www.theregister.co.uk/2014/09/13/gartner_ranks_startups_ahead_of_emc_hp_and_ibm/. Gartner produces two reports for two different audiences. The MQ reviews companies, that is vendors on their ability to succeed. The Critical Capabilities (CC) reviews vendors products on their suitability for workloads and the products features. There are links to them here http://blogs.gartner.com/valdis-filks/2014/08/29/solid-state-array-magic-quadrant-and-critical-capabilities-published/. The discussion here is more orientated to the technical aspects of solid state arrays, so you should really reference the CC, which does explain the points which you mention above. If you are a Gartner client please organise a call and we can have a chat.
Sunday 15th February 2015 23:39 GMT boriss111
True role of AFA
Before we start debate if built-from-scratch AFA or evolved-enough-to-be AFA storage is better, we really need to consider which role we have in mind for the all flash storage array. As everything else in life, there is always a trade-off with storage systems. Built-from-scratch AFA shine with their killer performance (measured by latency) under heavy workloads (measured by IOPS). But this comes with price. We cannot expect too much services to be offered as these would, without doubt, destroy performance. Maybe a bit extreme example would be AFA with synchronous replication over 100km distance running on top. What would be the latency? This instantly disqualify built-from-scratch AFA storage as general purpose storage and more importantly, being only storage in the datacentre. Additionally, extreme performance of built-from-scratch AFA is seldom needed for all workloads in a typical environment.
On the other hand, evolved-enough-to-be AFA doesn't offer that much performance under heavy workloads, but come at much more attractive price point. Sharing majority of components with a general purpose storage which is being produced in huge quantities have something to do with that. Being born out of disk based storage also brings full array of storage services which allow evolved-enough-to-be AFA to be only storage in environments which require high performance across the board.
Having all this in mind, we should consider that role of built-from-scratch AFA and evolved-enough-to-be AFA is not the same, therefore comparison between two is not exactly apple to apple comparison.
(disclosure: working for EMC partner company)
Friday 21st August 2015 00:14 GMT HPStorageGuy
HP 3PAR is number 2 AFA in the market
(Calvin Zito from HP Storage here)
Old article and the comments have been stale for a long time but I sure hope the AFA "Startups" have caught on to the fact that HP 3PAR is a fantastic architecture for flash. Lots of support for that in many of the comments above but I'm still amused by my colleagues in the industry who tried to make 3PAR irrelevant in all-flash because it wasn't built from scratch only for flash.
Given the recent success in the industry, seems as though the "you're not flash" argument has fallen on deaf ears. Specifically:
> 3PAR all-flash was named AFA of the year by Storage Magazine and won best of show at last week's Flash Memory Summit. Lots of other awards but I'll spare you that.
> HP 3PAR All-Flash was the fastest growing AFA in the market, moving from number 10 to number 5 in 2014 with over 1000% growth. Yup, we started from a small base but goes to show you the acceptance of 3PAR AFA in a short time.
> With the latest revelations that Pure Storage has overstated their revenue numbers, HP 3PAR AFA has taken the number 2 spot in all-flash revenue.
Since the built-from-scratch argument didn't work, now we're hearing competitive FUD about how using an ASIC is a poor architecture choice. Here's a post that puts that to rest: http://h30507.www3.hp.com/t5/Around-the-Storage-Block-Blog/Commodity-versus-Custom-Fear-the-ASIC/ba-p/184057
What a fun industry storage is, huh?