All Flash Arrays: Scale Up vs Scale Out (Part 1)

Imagine you want to buy some more storage for your laptop – let’s say an external USB drive for backups. What are the fundamental questions you need to ask before you get down to the thorny issue of price? Typically, there is only one key question:

  • How much capacity do I need?

Of course there will be lesser questions, such as connectivity, brand, colour, weight, what it actually looks like and so on. But those are qualifying questions – ways to filter the drop down list on Amazon so you have less decisions to make.

However, different rules apply when buying enterprise storage: we might care less about colour and more about physical density, power requirements and the support capabilities of the vendor. We might care less about what the product looks like and more about how simple it is to administer. But most of all, for enterprise storage, there are now two fundamental questions instead of one:

  • How much capacity do I need?
  • How much performance do I need?

Of course, there is the further issue of what exactly we mean by “performance”, given that it can be measured in a number of different ways. The answer is dependant on the platform being used: for disk-based storage systems it was typically the number of I/O Operations Per Second (IOPS), while for modern storage systems it is more likely to be the bandwidth (the volume of data read and written per second). And just to add a little spice, when considering IOPS or bandwidth on All Flash array platforms the read/write ratio is also important.

So the actual requirement for, say, a three-node data warehouse cluster with 200 users might turn out to be:

“I need 50TB of usable capacity and the ability to deliver 1GB/second at 90% reads. What will this cost?”

Are we ready to spec a solution yet? Not quite. First we have to consider Rule #1.

Rule #1: Requirements Change

Most enterprise storage customers purchase their hardware for use over a period of time – with the most typical period being five years. So it stands to reason that whatever your requirements are at the time of purchase, they will change before the platform is retired. In fact, they will change many times. graph-growingData volumes will grow, because data volumes only ever get bigger… right? But also, those 200 users might grow to 500 users. The three node cluster might be extended to six nodes. The chances are that, in some way, you will need more performance and/or more capacity.

The truth is that, while customers buy their hardware based on a five year period, now more than ever they cannot even predict what will happen over the next 12 months. Forgive the cliche, the only thing predictable is unpredictability.

So as a customer in need of enterprise storage, what do you do? Clearly you won’t want to purchase all of the capacity and performance you might possibly maybe need in the future. That would be a big up-front investment which may never achieve a return. So your best bet is to purchase a storage platform which can scale. This way you can start with what you need and scale as your requirements grow.

This is where architecture becomes key.

Scalability is an Architectural Decision

You may remember from a previous post that the basic building blocks of any storage array are controllers and media, with various networking devices used to string them together. scale-starting-pointAs a gross simplification, the performance of a storage array is a product of the number of controllers and the power that those controllers have (assuming they aren’t held back by the media). The capacity of a storage array is clearly a product of the amount of media.

In a simple world you would just add more media when you wanted more capacity, you would add more controllers (or increase their power) when you needed more performance, and you would do both when you need both. But this is not a simple world.

I always think that the best way to visualise the two requirements of capacity and performance is to use two different dimensions on a graph, with performance as X and capacity as Y. So let’s use that here – we start with a single flash array which has one pair of controllers and one set of flash media.

Scale Up – Adding Capacity

The simple way to achieve scale up is to just add more media to an existing array. Media is typically arranged in some sort of indivisible unit, such as a shelf of SSDs arranged in a RAID configuration (so that it has inbuilt redundancy). In principle, adding another shelf of SSDs sounds easy, but complications arise when you consider the thorny issue of metadata.

To illustrate, consider that most enterprise All Flash arrays available today have inbuilt data reduction features including deduplication. At a high level, dedupe works by computing a hash value for each block written to the array and then comparing it to a table containing all the hash values of previously written blocks. If the hash is discovered in the table, the block already exists on the array and does not need to be written again, reducing the amount of physical media used.

This hash table is an example of the metadata storage arrays need to store and maintain in order to function; other examples of features which utilise metadata are thin provisioning, snapshots and replication. This makes metadata a critical factor in the performance of a storage array

To ensure the highest speed of access, most metadata is pinned in DRAM on the storage controllers. This has a knock-on effect in that the amount of addressable storage in an array can become directly linked to the amount of DRAM in the controllers. DRAM is a costly resource and affects the manufacturing cost of the array, so there is a balancing act required in order to have enough DRAM to store the necessary metadata without inflating the cost any more than is absolutely possible.

Hopefully you see where this is going. Adding more shelves of media increases the storage capacity of an array… thereby increasing the metadata footprint… and so increasing the need for DRAM. At some point the issue becomes whether it is even possible (technically or commercially) to support the metadata overhead of adding more shelves of physical media.

Scale-Up Only Architectures

There are many storage arrays on the market which have a scale-up only architecture, with Pure Storage being an obvious example. There are various arguments presented as to why this is the case, but my view is that these architectures were used as a compromise in order to get the products to market faster (especially if they also adopt an active/passive architecture). Having said that, it’s obvious that I am biased by the fact that I work for vendor which does not have this restriction – and who believes in offering scale up and scale out. So please don’t take my word for it – go read what the other vendors say and then form your own opinion.

One counter claim by proponents of scale-up only architecture is that performance can be added by upgrading array controllers in-place, non-disruptively. In other words, the controllers can be replaced with high-specification models with more CPU cores and more DRAM, bringing more performance capability to the array. The issue here is that this is a case of diminishing returns. Moving up through the available CPU models brings step changes in cost but only incremental increases in performance.

To try and illustrate this, let’s look at some figures for the Pure Storage //m series of All Flash arrays. There are three models increasing in price and performance: m20, m50 and m70. We can get performance figures measured in maximum IOPS (measured with a 32k block size as is Pure’s preferred way) from this datasheet and we can get details of the CPU and DRAM specifications from this published validation report by the NSA. Let’s use Wikipedia’s List of Intel Xeon microprocessors page to find the list price of those CPUs and compare the increments in price to those of the maximum stated performance:

Here we can see that the list price for the CPUs alone rises 238% and then 484% moving from //m20 to //m50 to //m70, yet the maximum performance measured in IOPS rises just 147% and 200%. You can argue that the CPU price is not a perfect indicator of the selling price of each //m series array, but it’s certainly a factor. As anyone familiar with purchasing servers will attest, buying higher-spec models takes more out of your pocket than it gives you back in performance.

My point here – and this is a general observation rather than one about Pure in particular – is that this is not a cost effective scaling strategy in comparison to the alternative, which is the ability to scale out by adding more controllers.

Coming Next: Scale Out

In the next post we’ll look at Scale Out architectures and what they mean for customers with independently varying requirements for capacity and performance.

Scale out allows the cost-effective addition of more controllers and therefore more performance capability, along with other benefits such as the addition of more ports. But there are potential downsides too…