Patrick Allaire

Bed Time Stories for Flash Buyers

Blog Post created by Patrick Allaire Employee on Nov 3, 2016

AFA Sales Pitch vs Latent Weakness

 

I just finished another round of product training and internal communications for the Hitachi Storage Virtualization Operating System (SVOS) 7 release … I am now seeing the light at the end of the tunnel.

 

I wanted to share my take on the feedback I received over the last few weeks from sales, partners, and early customers.  This post should be of interest to all-flash array (AFA) buyers and sales professionals.

 

Let’s start with the usual disclaimers and apologies, I recognize that if you are a customer browsing through a flash vendor’s website or reading one of their brochures or whitepapers, you may feel that all the storage systems you have in production are old and in need of a refresh; similar to seeing an Apple announcement on their new iPhone, just minus the fancy video.

 

Bed Time stories.jpgNot surprisingly, flash vendors are messaging a story about simpler and easier-to-use technology with a few unique features positioned as “breakthrough innovations”.

 

Like the stories I tell my 3 years old son before bed time, the flash vendor stories all end well.  The hero, the all-flash array, always wins at the end.

 

Hitachi’s flash stories though are more like documentaries. Hitachi’s Japanese culture is so strong, built on honesty and open disclosures, that each time we do a new release we are pushed to be brutally honest. As an example, some Hitachi staff will give us grief about our literature unless every benefit is applicable to every use case and every environment.

 

In a way, these individuals are demanding that we call sushi: Cold.  Dead.  Fish, on rice. I apologize that we cannot comply with your request. We need our customers to see beyond the technical facts and understand the differentiated VALUE, that we bring to customers. I will promise you though that we will not over-hype common industry technologies like they were grand new things that have never been seen before. HDS does not tell bed time stories.

 

Mousetrap.png

To flash buyers, beware of AFA product promises. The benefits built on the promise of easy management, low cost through always-on data reduction and high speed, low latency are often much prettier than the reality of what they deliver. Especially after the demo has completed and you have used them in production for six months.

 

This is because many AFA sales pitches are built on comparison to hard disk drives (HDDs) storage arrays. That boat sailed a long time ago and is no longer a realistic comparison. It is like comparing ‘the little engine that could’ to Telsa. You need a better compare point.

Of course flash provides better latency and consistency of response time out of the box versus any HDDs. A more interesting comparison is how these AFAs perform when compare against each other.

 

 

I have seen good papers from analysts on flash TCO and cost reduction attributed to lower management overhead, better environmental characteristics and reduced host licensing costs; I see far less public disclosure from big petabyte flash buyers who invest the resources and time to compare AFAs value.

Headinthesand.pngWhile the Storage Performance Council (SPC) was a great value for traditional storage comparison, no AFA vendors as yet elected to published any SPC-1 benchmark results using version 3.2 specifications that support data reduction.  Prior flash storage system SPC-1 results have limited value in AFA comparison since none of the data reduction features were “in use” with these benchmarks.

 

When you look at the validation efforts behind our largest buyer’s proof of concepts; it becomes clear that the average flash buyer is at a disadvantage. With no all-flash with data reduction SPC-1 results to use for comparing products, designing a custom benchmark is labor intensive and require analysis over a significant period of time to see how the arrays perform AFTER the honeymoon period. The period when regular AFA tasks start to run and latencies go up. Drastically up for certain vendors.

 

 

Flying.pngWhile, I could write this entire blog post on primary workload profiling and how long you have to run tests to determine real performance and latency, I will limit my comments the the following: push the vendors and do not be that guy - the one that improvises. Improvising any benchmark without some planning is risky at best and a career limiting move at worst. What is the value of testing, if the results will not look the same after 90 days of deployment? If you still want to pursue this route, the best advice I’ve seen was from an analyst paper is from: IDC report on All-Flash Array Performance Testing Framework by Dan Iacono.  Avoid testing small block, read-only workload as this profile is not representative of any primary workload.

While every workload is different, the common agreement between competing flash vendors is that primary workload average block size tends to be much larger; 32KB or more.  Testing a 4KB or 8KB workload block size is like doing a test drive of a sports car in the dealership parking lot.

 

 

I made the same point during HDS ambassador training for the SVOS 7 release.  Using results from a recent, enterprise customer proof of concept I demonstrated the difference between AFA demos (the honeymoon phase) and deployment experiences (where the rubber meets the road).  For your reference, those results are shared here. I took out the competitor vendor and model information to prevent vendor bashing while illustrating how the impact of data reduction changes the AFA experience over time.

Response time impact.png

While AFA vendors and large flash buyer aren’t surprised by the disparity of results, the typical AFA buyer is often not aware that his flash investment (model A – E) will underperform over time and the bed time story becomes more of a nightmare.This is not new, every storage array is limited by its CPU and memory resources. It is just that data reduction services are just taxing the system even more, often to the point where applications experience  less than stellar response times…

 

What can you do about it as a flash buyer?  Do not buy based on AFA brand hype or specifications alone. Some of the top AFA vendors are models A – E.  Instead, leverage your peer network to get insights from actual deployment – hopefully with your specific workload.  Also, seek out AFAs that support selective data reduction services that can be turned on or off. This enables you to make the call about when and where they are used. It empowers you to make decisions on when the value of capacity savings outweighs the performance overhead from theses data services.

 

What will you do with your extra capacity?  I am looking forward to your feedback.

 

For the curious reader, call your HDS sales representative….  They know who is behind AFA model A to E.  By the way, in this proof of concept, Hitachi was brand A

 

 

I hope you had fun at Halloween, I surely did.

Hallowen with family.png

Outcomes