It has been an eventful week in the storage industry with lots of all-flash array (AFA) news - and it hasn't been good for several big vendors. Over the last couple of years as I've been working with our HP 3PAR team to get my head around AFAs, the thing I've heard repeatedly and consistently is that architecture matters.
By Calvin Zito, @HPStorageGuy
So I want to do a couple of things in our 3PAR Thursday blog post today - share some of that news that makes it clear that architecture does matter and I have a video from our lead HP 3PAR Software Product Manager Ivan Iannaccone talking about why flash architecture matters with lead analyst from Storage Switzerland George Crump.
First, the news
It really is hard to believe some of the news. First, Cisco announced that they stopped Invicta shipments due to a "scalability issue". Invicta is their all-flash acquisition of Whiptail. And just yesterday, NetApp announced that they'll start shipping their long awaited AFA, FlashRay, with - get this - a single controller. Chris Mellor from the Register said it best, "It's not ready for prime time, strangely." It's designed from the ground up but is being released with limited scale and one controller. Architecture matters!
Then earlier this week, news broke in a blog by Andrew Dauncey that EMC XtremIO upgrades from 2.4 to 3.0 would be a distruptive software upgrade. Basically, customers will have to back up their data, the array will be wiped during the upgrade and you'll need to rebuild your array. Wow! The question customers should be asking EMC is why would you design a product for mission-critical applications that can't support non-disruptive upgrades (NDU)?
My guess is that the disruptive upgrade is required because EMC is changing their dedupe block size from 4K to 8K. Using 4K requires a lot of metadata which on XtremIO is 100% managed inside controller cache. Controller cache is limited and that limits the scalability of the system. And that is a competitive disadvantage. Metadata management is a fundamental storage architecture decision and EMC didn't get it right before actually shipping ExtremIO. Architecture matters!
Blogger Chris Evans has a very good blog post that also discusses this that you should check out.
Here's a graphic that Ivan gave me that speaks to scale out, comparing dual controller AFAs with 3PAR.
One comment about active/passive architectures: be sure to understand what you're getting there. One AFA markets it as a benefit that "no performance impact when a controller fails." That's fine but that means that 99% of the time that both controllers are avalable, one of them is not helping get the most performace out of your system. Architecture matters.
End-to-End Virtualization and I/O Optimization
Another interesting differentiation that the HP 3PAR StoreServ solution brings is the architecture's ability to cover all aspects of the life of an I/O. While other solutions concentrate mainly on optimization within the controller and how data is written to flash, HP 3PAR StoreServ approach is to optimize the entire I/O stack from how the host I/O is wide striped and prioritized down to how this is eventually written to flash. Here's another image that Ivan gave me that helps with this.
Video with Storage Switzerland and Ivan
One last thought - Pure Storage jumped into this (no surprise) and said they've never had a NDU, except for a couple times when they rolled customers from beta software to general availability. That's good but we don't think it's good enough. In 12 years of HP 3PAR, we've never required a data destructive upgrade including when we introduced HP 3PAR AFA. Architecture matters.