Architecture has long term scalability implications for All Flash Appliances

Recently, NetApp announced the availability of a 3.84TB SSD. It’s not extremely exciting – it’s just a larger storage medium. Sure, it’s really advanced 3D NAND, it’s fast and ultra-reliable, and will allow some nicely dense configurations at a reduced $/GB. Another day in Enterprise Storage Land.

But, ultimately, that’s how drives roll – they get bigger. And in the case of SSD, the roadmaps seem extremely aggressive regarding capacities.

Then I realized that several of our competitors don’t have this large SSD capacity available. Some don’t even have half that.

But why? Why ignore such a seemingly easy and hugely cost-effective way to increase density?

In this post I will attempt to explain why certain architectural decisions may lead to inflexible design constructs that can have long-term flexibility and scalability ramifications.

Design Center

Each product has its genesis somewhere. It is designed to address certain key requirements in specific markets and behave in a better/different way than competitors in some areas. Plug specific gaps. Possibly fill a niche or even become a new product category.

This is called the “Design Center” of the product.

Design centers can evolve over time. But, ultimately, every product’s Design Center is an exercise in compromise and is one of the least malleable parts of the solution.

There’s no such thing as a free lunch. Every design decision has tradeoffs. Often, those tradeoffs sacrifice long term viability for speed to market. There’s nothing wrong with tradeoffs as long as you know what those are, especially if the tradeoffs have a direct impact on your data management capabilities long term.

It’s all about the Deduplication/RAM relationship

Aside from compression, scale up and/or scale out, deduplication is a common way to achieve better scalability and efficiencies out of storage.

There are several ways to offer deduplication in storage arrays: Inline, post-process, fixed chunk, variable chunk, per volume scope, global scope – all are design decisions that can have significant ramifications down the line and various business impacts.

Some deduplication approaches require very large amounts of memory to store metadata (hashes representing unique chunk signatures). This may limit scalability or make a product more expensive, even with scale-out approaches (since many large, costly controllers would be required).

There is no perfect answer, since each kind of architecture is better at certain things than others. This is what is meant by “tradeoffs” in specific Design Centers. But let’s see how things look for some example approaches (this is not meant to be a comprehensive list of all permutations).

I am keeping it simple – I’m not showing how metadata might get shared and compared between nodes (in itself a potentially hugely impactful operation as some scale-out AFA vendors have found to their chagrin). In addition, I’m not exploring container vs global deduplication or different scale-out methods – this post would become unwieldy… If there’s interest drop me a line or comment and I will do a multi-part series covering the other aspects.

Fixed size chunk approach

In the picture below you can see the basic layout of a fixed size chunk deduplication architecture. Each data chunk is represented by a hash value in RAM. Incoming new chunks are compared to the RAM hash store in order to determine where and whether they may be stored:

Hashes fixed chunk

The benefit of this kind of approach is that it’s relatively straightforward from a coding standpoint, and it probably made a whole lot of sense a couple of years ago when small SSDs were all that was available and speed to market was a major design decision.

The tradeoff is that a truly exorbitant amount of memory is required in order to store all the hash metadata values in RAM. As SSD capacities increase, the linear relationship of SSD size vs RAM size results in controllers with multi-TB RAM implementations – which gets expensive.

It follows that systems using this type of approach will find it increasingly difficult (if not impossible) to use significantly larger SSDs without either a major architectural change or the cost of multiple TB of RAM dropping dramatically. You should really ask the vendor what their roadmap is for things like 10+TB SSDs… and whether you can expand by adding the larger SSDs into a current system without having to throw everything you’ve already purchased away.

Variable size chunk approach

This one is almost identical to the previous example, but instead of a small, fixed block, the architecture allows for variable size blocks to be represented by the same hash size:

Hashes variable chunk

This one is more complex to code, but the massive benefit is that metadata space is hugely optimized since much larger data chunks are represented by the same hash size as smaller data chunks. The system does this chunk division automatically. Less hashes are needed with this approach, leading to better utilization of memory.

Such an architecture needs far less memory than the previous example. However, it is still plagued by the same fundamental scaling problem – only at a far smaller scale. Conversely, it allows a less expensive system to be manufactured than in the previous example since less RAM is needed for the same amount of storage. By combining multiple inexpensive systems via scale-out, significant capacity at scale can be achieved at a lesser cost than with the previous example.

Fixed chunk, metadata both in RAM and on-disk

An approach to lower the dependency on RAM is to have some metadata in RAM and some on SSD:

Hashes fixed chunk metadata on disk

This type of architecture finds it harder to do full speed inline deduplication since not all metadata is in RAM. However, it also offers a more economical way to approach hash storage. SSD size is not a concern with this type of approach. In addition, being able to keep dedupe metadata on cold storage aids in data portability and media independence, including storing data in the cloud.

Variable chunk, multi-tier metadata store

Now that you’ve seen examples of various approaches, it starts making logical sense what kind of architectural compromises are necessary to achieve both high deduplication performance and capacity scale.

For instance, how about variable blocks and the ability to store metadata on multiple tiers of storage? Upcoming, ultra-fast Storage Class Memory technologies are a good intermediate step between RAM and SSD. Lots of metadata can be placed there yet retain high speeds:

Hashes variable chunk metadata on 3 tiers

Coding for this approach is of course complex since SCM and SSD have to be treated as a sort of Level 2/Level 3 cache combination but with cache access time spans in the days or weeks, and parts of the cache never going “cold”. It’s algorithmically more involved, plus relies on technologies not yet widely available… but it does solve multiple problems at once. One could of course use just SCM for the entire metadata store and simplify implementation, but that would somewhat reduce the performance afforded by the approach shown (RAM is still faster). But if the SCM is fast enough… 🙂

However, being able to embed dedupe metadata in cold storage can still help with data mobility and being able to retain deduplication even across different types of storage and even cloud. This type of flexibility is valuable.

Why should you care?

Aside from the academic interest and nerd appeal, different architecture approaches have a business impact:

  • Will the storage system scale large enough for significant future growth?
  • Will I be able to use significantly new media technologies and sizes without major disruption?
  • Can I use extremely large media sizes?
  • Can I mix media sizes?
  • Can I mix controller types in a scale-out cluster?
  • Can I use cost-optimized hardware?
  • Does deduplication at scale impact performance negatively, especially with heavy writes?
  • If inline efficiencies aren’t comprehensive, how does that affect overall capacity sizing?
  • Does the deduplication method enforce a single large failure domain? (single pool – meaning that any corruption would result in the entire system being unusable)
  • What is the interoperability with Cloud and Disk technologies?
  • Can data mobility from All Flash to Disk to Cloud retain deduplication savings?
  • What other tradeoffs is this shiny new technology going to impose now and in the future? Ask to see a 5-year vision roadmap!

Always look beyond the shiny feature and think of the business benefits/risks. Some of the above may be OK for you. Some others – not so much.

There’s no free lunch.

D

Technorati Tags: , , , ,

NetApp posts SPC-1 Top Ten Performance results for its high end systems – Tier 1 meets high functionality and high performance

It’s been a while since our last SPC-1 benchmark submission with high-end systems in 2012. Since then we launched all new systems, and went from ONTAP 8.1 to ONTAP 8.3, big jumps in both hardware and software.

In 2012 we posted an SPC-1 result with a 6-node FAS6240 cluster  – not our biggest system at the time but we felt it was more representative of a realistic solution and used a hybrid configuration (spinning disks boosted by flash caching technology). It still got the best overall balance of low latency (Average Response Time or ART in SPC-1 parlance, to be used from now on), high SPC-1 IOPS, price, scalability, data resiliency and functionality compared to all other spinning disk systems at the time.

Today (April 22, 2015) we published SPC-1 results with an 8-node all-flash high-end FAS8080 cluster to illustrate the performance of the largest current NetApp FAS systems in this industry-standard benchmark.

For the impatient…

  • The NetApp All-Flash FAS8080 SPC-1 submission places the system in the #5 performance spot in the SPC-1 Top Ten by performance list.
  • And #3 if you look at performance at load points around 1ms Average Response Time (ART).
  • The NetApp system uses RAID-DP, similar to RAID-6, whereas the other entries use RAID-10 (typically, RAID-6 is considered slower than RAID-10).
  • In addition, the FAS8080 shows the best storage efficiency, by far, of any Top Ten SPC-1 submission (and without using compression or deduplication).
  • The FAS8080 offers far more functionality than any other system in the list.

We also recently posted results with the NetApp EF560  – the other major hardware platform NetApp offers. See my post here and the official results here. Different value proposition for that platform – less features but very low ART and great cost effectiveness are the key themes for the EF560.

In this post I want to explain the current Clustered Data ONTAP results and why they are important.

Flash performance without compromise

Solid state storage technologies are becoming increasingly popular.

The challenge with flash offerings from most vendors is that customers typically either have to give up a lot in order to get the high performance of flash, or have to combine 4-5 different products into a complex “solution” in order to satisfy different requirements.

For instance, dedicated all-flash offerings may not be able to natively replicate to less expensive, spinning-drive solutions.

Or, a flash system may offer high performance but not the functionality, scalability, reliability and data integrity of more mature solutions.

But what if you could have it all? Performance and reliability and functionality and scalability and maturity? That’s exactly what Clustered Data ONTAP 8.3 provides.

Here are some Clustered Data ONTAP 8.3 running on FAS8080 highlights:

  • All the NetApp signature ultra-tight application integration and automation for replication, SnapShots, Clones
  • Fancy write-accelerated RAID6-equivalent protection by default
  • Comprehensive data integrity and protection against insidious lost write/torn page/misplaced write errors that RAID and normal checksums don’t always catch
  • Non-disruptive data mobility for all protocols
  • Non-disruptive operations – no downtime even when doing things that would require downtime and extensive PS with other vendors
  • Granular QoS
  • Deduplication and compression
  • Highly scalable – 5,760 drives possible in an 8-node cluster, 17,280 drives possible in the max 24 nodes. Various drive types in the cluster, from SSD to SATA and everything else in between.
  • Multiprotocol (FC, iSCSI, NFS, SMB1,2,3) on the same hardware (no “helper” boxes needed, no dedicated SAN vs NAS pools needed)
  • 96,000 LUNs per 8-node cluster (that’s right, ninety-six thousand LUNs – about 50% more than the maximum possible with the other high-end systems)
  • ONTAP is VMware vVol ready
  • The only array that has been validated by VMware for VMware Horizon 6 with vVols – hopefully the competitors will follow our lead
  • Over 460TB (yes, TeraBytes) of usable cache after all overheads are accounted for (and without accounting for cache amplification through deduplication and clones) in an 8-node cluster. Makes competitor maximum cache amounts seem like rounding errors – indeed, the actual figure might be 465TB or more, but it’s OK… 🙂 (and 3x that number in a 24-node cluster, over 1.3PB cache!)
  • The ability to virtualize other storage arrays behind it
  • The ability to have a cluster with dissimilar size and type nodes – no need to keep all engines the same (unlike monolithic offerings). Why pay the same for all nodes when some nodes may not need all the performance? Why be forced to keep all nodes in the same hardware family? What if you don’t want to buy all at once? Maybe you want to upgrade part of the cluster with a newer-gen system? 🙂
  • The ability to evacuate part of a cluster and build that part as a different cluster elsewhere
  • The ability to have multiple disk types in a cluster and, indeed, dedicate nodes to functions (for instance, have a few nodes all-flash, some nodes with flash-accelerated SAS and a couple with very dense yet flash-accelerated NL-SAS, with full online data mobility between nodes)
That last bullet deserves a picture:
 MixedCluster

 

“SVM” stands for Storage Virtual Machine –  it means a logical storage partition that can span one or more cluster nodes and have parts of the underlying capacity (performance and space) available to it, with its own users, capacity and performance limits etc.

In essence, Clustered Data ONTAP offers the best combination of performance, scalability, reliability, maturity and features of any storage system extant as of this writing. Indeed – look at some of the capabilities like maximum cache and number of LUNs. This is designed to be the cornerstone of a datacenter.

it makes most other systems seem like toys in comparison…

Ships

FUD buster

Another reason we wanted to show this result was FUD from competitors struggling to find an angle to fight NetApp. It goes a bit like this: “NetApp FAS systems aren’t real SAN, it’s all simulated and performance will be slow!”

Right…

Drevilsimulated

Well – for a “simulated” SAN (whatever that means), the performance is pretty amazing given the level of protection used (RAID6-equivalent – far more resilient and capacity-efficient for large pooled deployments than the RAID10 the other submissions use) and all the insane scalability, reliability and functionality on tap 🙂

Another piece of FUD has been that ONTAP isn’t “flash-optimized” since it’s a very mature storage OS and wasn’t written “from the ground up for flash”. We’ll let the numbers speak for themselves. It’s worth noting that we have been incorporating a lot of flash-related innovations into FAS systems well before any other competitor did so, something conveniently ignored by the FUD-mongers. In addition, ONTAP 8.3 has a plethora of flash optimizations and path length improvements that helped with the excellent response time results. And lots more is coming.

The final piece of FUD we made sure was addressed was system fullness – last time we ran the test we didn’t fill up as much as we could have, which prompted the FUD-mongers to say that FAS systems need gigantic amounts of free space to perform. Let’s see what they’ll come up with this time 😉

On to the numbers!

As a refresher, you may want to read past SPC-1 posts here and here, and my performance primer here.

Important note: SPC-1 is a 100% block-based benchmark with its own I/O blend and, as such, the results from any vendor SPC-1 submission should not be compared to marketing IOPS numbers of all reads or metadata-heavy NAS benchmarks like SPEC SFS (which are far easier on systems than the 60% write blend of the SPC-1 workload). Indeed, the tested configuration might perform in the millions of “marketing” IOPS – but that’s decidedly not the point of this benchmark.

The SPC-1 Result links if you want the detail are here (summary) and here (full disclosure). In addition, here’s the link to the “Top 10 Performance” systems page so you can compare other submissions that are in the upper performance echelon (unfortunately, SPC-1 results are normally just alphabetically listed, making it time-consuming to compare systems unless you’re looking at the already sorted Top 10 list).

I recommend you look beyond the initial table in each submission showing the performance and $/SPC-1 IOPS and at least go to the price table to see the detail. The submissions calculate $/SPC-1 IOPS based on submitted price but not all vendors use discounted pricing. You may want to do your own price/performance calculations.

The things to look for in SPC-1 submissions

Typically you’re looking for the following things to make sense of an SPC-1 submission:

  • ART vs IOPS – many submissions will show high IOPS at huge ART, which would be rather useless when it comes to Flash storage
  • Sustainability – was performance even or are there constant huge spikes?
  • RAID level – most submissions use RAID10 for speed, what would happen with RAID6?
  • Application Utilization. This one is important yet glossed over. It signifies how much capacity the benchmark consumed vs the overall raw capacity of the system, before RAID, spares etc.

Let’s go over these one by one.

ART vs IOPS

Our ART was 1.23ms at 685,281.71 SPC-1 IOPS, and pretty flat over time during the test:

Response_time_complete

Sustainability

The SPC-1 rules state the minimum runtime should be 8 hours. We ran the test for 18 hours to observe if there would be variation in the performance. There was no significant variation:

IOdistributionRamp

RAID level

RAID-DP was used for all FAS8080EX testing. This is mathematically analogous in protection to RAID-6. Given that these systems are typically deployed in very large pooled configurations, we elected long ago to not recommend single parity RAID since it’s simply not safe enough. RAID-10 is fast and fine for smaller capacity SSD systems but, at scale, it gets too expensive for anything but a lab queen (a system that nobody in their right mind will ever buy but which benchmarks well).

Application Utilization

Our Application Utilization was a very high 61.92% – unheard of by other vendors posting SPC-1 results since they use RAID10 which, by definition, wastes half the capacity (plus spares and other overheads to worry about on top of that).

AppUtilization

Some vendors using RAID10 will fill up the resulting space after RAID, spares etc. to a very high degree, and call out the “Protected Application Utilization” as being the key thing to focus on.

This could not be further from the truth – Application Utilization is the only metric that really shows how much of the total possible raw capacity the benchmark actually used and signifies how space-efficient the storage was.

Otherwise, someone could do quadruple mirroring of 100TB, fill up the resulting 25TB to 100%, and call that 100% efficient… when in fact it only consumed 25% 🙂

It is important to note there was no compression or deduplication enabled by any vendor since it is not allowed by the current version of the benchmark.

Compared to other vendors

I wanted to show a comparison between the Top Ten Performance results both in absolute terms and also normalized around 1ms ART.

Here are the Top Ten highest performing systems as of April 22, 2015, with vendor results links if you want to look at things in detail:

FYI, the HP XP 9500 and the Hitachi system above it in the list are the exact same system, HP resells the HDS array as their high-end offering.

I will show columns that explain the results of each vendor around 1ms. Why 1ms and not more or less? Because in the Top Ten SPC-1 performance list, most results show fairly low ART, but some have very high ART, and it’s useful to show performance at that lower ART load point, which is becoming the ART standard for All-Flash systems. 1ms seems to be a good point for multi-function SSD systems (vs simpler, smaller but more speed-optimized architectures like the NetApp EF560).

The way you determine the 1ms ART load point is by looking at the table that shows ART vs SPC-1 IOPS. Let’s pick IBM’s 780 since it has a very interesting curve so you learn what to look for.

From page 5 of the IBM Power Server 780 SPC-1 Executive Summary:

IBM780

IBM’s submitted SPC-1 IOPS are high but at a huge ART number for an all-SSD solution (18.90ms). Not very useful for customers picking an all-SSD system. Even the next load point, with an average ART of 6.41ms, is high for an all-flash solution.

To more accurately compare this to the rest of the vendors with decent ART, you need to look at the table to find the closest load point around 1ms (which, in this case, it’s the 10% load point at 0.71ms – the next one up is much higher at 2.65ms).

You can do a similar exercise for the rest, it’s worth a look – I don’t want to paste all these tables and graphs since this post will get too big. But it’s interesting to see how SPC-1 IOPS vs ART are related and translate that to your business requirements for application latency.

Here’s the table with the current Top Ten SPC-1 Performance results as of 4/22/2015. Click on it for a clearer picture, there’s a lot going on.

8080_chart_spc1

Key for the chart (the non-obvious parts anyway):

  • The “SPC-1 Load Level near 1ms” is the load point in each SPC-1 Report that corresponds to the SPC-1 IOPS achieved near 1ms. This is not how busy each array was (I see this misinterpreted all the time).
  • The “Total ASU Capacity” is the amount of capacity the test consumed.
  • The “Physical Storage Capacity” is the total amount of capacity in the array before RAID etc.
  • “Application Utilization” is ASU Capacity divided by Physical Storage Capacity.

What do the results show?

Predictably, all-flash systems trump disk-based and hybrid systems for performance and can offer very nice $/SPC-1 IOPS numbers. That is the major allure of flash – high performance density.

Some takeaways from the comparison:

  • Based on SPC-1 IOPs around 1ms Average Response Time load points, the FAS8080 EX shifts from 5th place to 3rd
  • The other vendors used RAID10 – NetApp used RAID-DP (similar to RAID6 in protection). What would happen to their results if they switched to RAID6 to provide a similar level of protection and efficiency?
  • Aside from the NetApp FAS result, the rest of the Top Ten Performance submissions offer vastly lower Application Utilization – about half! Which means that NetApp is able to use 2x the capacity vs raw compared to the other submissions. And that’s before starting to count the possible storage efficiencies we can turn on like dedupe and compression.

How does one pick a flash array?

It depends. What are you trying to do? Solve a tactical problem? Just need a lot of extra speed and far lower latency for some workloads? No need for the array to have a ton of functionality? A lot of the data management happens in the application? Need something cost-effective, simple yet reliable? Then an all-flash system like the NetApp EF560 is a solid answer, and it can still be front-ended by a Clustered Data ONTAP system to provide more functionality if the need arises in the future (we are firm believers in hardware reuse and investment protection – you see, some companies talk about Software Defined Storage, we do Software Defined Storage).

On the other hand, if you would prefer an Enterprise architecture that can serve as the cornerstone of your datacenter for almost any workload and protocol, offers rich data management functionality and tight application integration, insane scalability, non-disruptive everything and offers the most features (reliably) compared to any other platform – then the FAS line running Clustered Data ONTAP is the only possible answer.

Couple that with OnCommand Insight – the best multivendor fabric management tool on the planet – plus Workflow Automation, and we’ve got you covered.

In summary – the all-flash FAS8080EX gets a pretty amazing performance and efficiency SPC-1 result, especially given the extensive portfolio of functionality it offers. In my opinion, no competitor system offers the sheer functionality the FAS8080 does – not even close. Additionally, I believe that certain competitors have very questionable viability and/or tiny market penetration, making them a risky proposition for a high end system purchase.

Thx

D

Technorati Tags: , , , , , , , ,