You Can’t Have Enough Performance, Really.

Posted by Tim Stoakes on Feb 10, 2015 6:18:06 PM
Tim Stoakes
Find me on:

The topic for this episode of the Architecture Matters blog series is performance, and specifically why performance is something you can’t have enough of.


When Violin pioneered the all-flash array, our goal was to deliver a product and architecture that would be built from the ground up to deliver flash storage at amazing speeds.

The need for speed existed; customers were telling us they could not scale their mission-critical enterprise systems due to the inherent latency and IOPS issues of disk, garnering significant business impact.

With our products, we have helped customers dramatically improve their business. For instance, cutting inventory cycles from days to hours, compiling software code from hours to minutes or generating real-time mobile billing records immediately instead of hours after the billing cycle.

Fast forward to today: customers are now adopting flash for tier-1, as the cost of flash has come down to that of enterprise disk and all-flash arrays now offer the same features that many customers have been accustomed to.

But with the deployment of flash in tier-1 storage, there is still a misconception that performance matters less compared to some of the earlier tier-0 deployments.

At Violin, we could not disagree more with such a statement.

While it is true that, performance requirements per workload are less for a tier-1 type of workload than for a tier-0 one, the reality of the matter is that for tier-1, the number of workloads dramatically increases from a few to hundreds on an array.

This is compounded with the need to consolidate storage platforms and build converged rack infrastructures for cloud architectures, putting even more stress on the ability to support different type of workloads, each with its own performance and cost per GB requirements.

What this means is that performance is more important than ever when delivering flash performance for the entire data center.

This is also the reason that at Violin, we believe strongly in the next evolution of more intelligent data-reduction: instead of having inline data-reduction always on for each workload, we believe in offering the option to turn it off for specific workloads that need optimal performance.

What this allows is not only the consolidation of several tier-1 disk storage platforms into a single 3U array, but also the consolidation of tier-0, tier-1 and even tier-2 storage into that same highly integrated and compact platform.

Another aspect of performance that is mis-understood: the real significance of both IOPS and latency.

The number of IOPS is really important: Violin not only shipped the first all-flash array, but also the first all-flash array that delivered 1 million IOPS. But driving a high number of IOPS is only half of how we look at performance in Violin.

The other half, which completes the performance equation, is clearly latency. Latency has a direct impact to the responsiveness of the application and is the biggest challenge that cannot be inherently overcome by disk.

The experience of going from tens of milliseconds to hundreds of microseconds in terms of storage latency is simply astounding. This drop in latency eliminates all inefficiency and under-utilization of assets in the entire data center. The result is that many Violin customers end up taking out a significant number of their application servers, freeing up rack-space and reducing software license, power and cooling costs.

Now the combination of latency and IOPS is what matters most: a tier-1 all-flash storage platform that delivers a combination of ultra-low latency with very high IOPS implicitly means that the system is able to support many IO’s in parallel, making it the best choice to consolidate hundreds of workloads and truly transform the data center.

As I mentioned in earlier blogs, Violin’s mission has been to unlock all the performance benefits of flash and re-think storage from the ground up for flash. This is why we don’t use SSDs; we have the best hardware-based system-level garbage collection capabilities in the industry and the capability to drive the highest possible sustained IOPS at the lowest latency.

The key word is sustained: every all-flash array can deliver high peak performance (either initially or when not fully loaded with data), but where things get interesting is when the all-flash array gets to write a lot of data: this is where the peak performance drops to sustained performance due to garbage collection kicking in.

The underlying all-flash architecture at this point makes a world of difference between an array that’s built up from the ground and keeps delivering sustained massive amounts of IOPS at ultra-low latency, or an array based on SSD’s that only delivers a fraction of its peak performance.

With years of experience in building all-flash arrays, this is the only real way of comparing performance: sustained lowest latency at highest IOPS at any read or write mix combination. What’s more, we wake up every day to beat our own records.

If you would like to experience the Violin effect, please contact us and we would love to help you unlock all the potential of your business.

I’m keeping the topic of the next episode a secret, as I want it to be a surprise. Here’s a hint: I have now covered all three areas of what we think the right all-flash array architecture consists of, so you may be wondering which platform has it all? Stay tuned to find out!