All-flash Array: Storage Gold?

All-flash storage array (AFA) provides two major benefits for the data center. First, AFA enables capacity efficiency with consistent performance and a reduced storage footprint. Second, AFA usually includes a software overlay, which abstracts storage hardware functions into the software. Think software-defined storage. These features include deduplication, or the elimination of duplicate copies of data, data compression, and thin provisioning.

These two qualities combine to form a dynamic duo of awesomeness for infrastructure teams looking to optimize their applications and maximize the utility of their storage arrays. All-flash storage essentially optimizes CPU utilization, so the number of IOs per second (IOPs) per host increases and also reduces the number of host servers needed to service the IOs.

So all that glitters is gold, right? Not so fast. The figure below shows how AFA affects the data ecosystem. In the past, traditional storage performance was measured in terms of the number of IOPs and the most influential variable is the number of spindles. With AFA, spindle count doesn’t matter, so performance centers on average latency, and that latency is influenced by the number of applications that will be piled onto the AFA. This means the bottleneck moves from spindle count to hitting the storage capacity limit as well as running hot in the other subsystems in the overall application stack.

trad vs afa.png

Are you considering AFA in your data center environment? Have you already implemented AFA? What issues have you run into, if any? Let me know in the comments below.

And don't forget to join SolarWinds and Pure Storage as we examine AFA beyond the IOPs to highlight performance essentials and uptime during a live webcast on June 8th at 2PM EST.

  • I certainly don't disagree that not all production workloads require SSD.  With that being said, every workload we have moved to SSD has performed better.  The other significant factor for us is the manageability, the more different storage systems and different tier of storage systems we have to manage the more people hours we spend and people are more expensive than a few additional SSD systems.  As a service provider we manage a LOT of storage systems so having clean cut lines to determine what goes where and by limiting what we manage we see a lot of efficiency gains.

    Oh, and I drive a truck!  emoticons_grin.png

  • I think that is the missing piece of puzzle here -- do you understand your workloads?  If the answer is No then spend some time figuring that out first.  Archived data on SSD is insane.  Heck, even low IOPS workloads on SSD is insane.  You wouldn't by a truck if all you needed was a compact car.

    Oh wait -- maybe that's a bad example emoticons_grin.png

  • We are in the process of making the move to flash storage for pretty much all production workloads.  It still makes a lot of sense to have large spindle based systems for any type of archiving.

  • Getting cheaper but still not disk price.

  • It's been a while since I broke out the ol' Performance Monitor... Let the distraction commence!

    Over a 10 minute (non-maintenance) window today, we averaged 2300 writes/sec and saw a maximum around 29000 writes/secemoticons_shocked.png  Three zeroes on that 29; some pretty stout IOPS right there.  Average write latency was almost nothing, and the maximum latency seen in the window was .008 sec/write.  Not bad at all.

    That said, we're still seeing about 17 minutes of WRITELOG wait in a similar 10 minute window - don't forget how many transactions jbiggley said we doing every second (above).  And, we failed to appropriately setup our SSDs for optimal DB performance; only one array for all DB volumes.  So we're going to add two more SSDs to serve as a dedicated log drive.  Once complete, we'll have a relatively optimal drive config.  I'll post some new numbers after the change.

    Incidentally, DPA could use some IOPS numbers.  Extremely relevant to highly transactional environments; like, ummm... SolarWinds.  emoticons_wink.png

Thwack - Symbolize TM, R, and C