It has been 2 months into my new role at DDN as a Solutions Architect. With many revolving doors around me, I have been trying to find the essence, the critical cog of the data infrastructure that supports the accelerated computing of the Nvidia GPU clusters. The more I read and engage, a pattern emerged. I found that cog in the supercharged data paths between the storage infrastructure systems and the GPU clusters. I will share more.
To set the context, let me start with a wonderful article I read in CIO.com back in July 2024. It was titled “Storage: The unsung hero of AI deployments“. It was music to my ears because as a long-time practitioner in the storage technology industry, it is time the storage industry gets its credit it deserves.
What is the data path?
To put it simply, a Data Path, from a storage context, is the communication route taken by the data bits between the compute system’s processing and program memory and the storage subsystem. The links and the established sessions can be within the system components such as the PCIe bus or external to the system through the shared networking infrastructure.
High speed accelerated data paths
In the world of accelerated computing such as AI and HPC, there are additional, more advanced technologies to create even faster delivery of the data bits. This is the accelerated data paths between the compute nodes and the storage subsystems. Following on, I share a few of these technologies that are lesser used in the enterprise storage segment.
I am strong believer of using the right tool to do the job right. I have said this before 2 years ago, in my blog “Stating the case for a Storage Appliance approach“. It was written when I was previously working for an open source storage company. And I am an advocate of the crafter versus assembler mindset, especially in the enterprise and high- performance storage technology segments.
I have joined DDN. Even with DDN that same mindset does not change a bit. I have been saying all along that the storage appliance model should always be the mindset for the businesses’ peace-of-mind.
My view of the storage appliance model began almost 25 years. I came into NAS systems world via Sun Microsystems®. Sun was famous for running NFS servers on general Sun Solaris servers. NFS services on Unix systems. Back then, I remember arguing with one of the Sun distributors about the tenets of running NFS over 100Mbit/sec Ethernet on Sun servers. I was drinking Sun’s Kool-Aid big time.
When I joined Network Appliance® (now NetApp®) in 2000, my worldview of putting software on general purpose servers changed. Network Appliance®, had one product family, the FAS700 (720, 740, 760) family. All NetApp® did was to serve NFS services in the beginning. They were the NAS filers and nothing else.
I was completed sold on the appliance way with NetApp®. Firstly, it was my very first time knowing such network storage services could be provisioned with an appliance concept. This was different from Sun. I was used to managing NFS exports on a Sun SPARCstation 20 to Unix clients in the network.
Secondly, my mindset began to shape that “you have to have the right tool to the job correctly and extremely well“. Well, the toaster toasts bread very well and nothing else. And the fridge (an analogy used by Dave Hitz, I think) does what it does very well too. That is what the appliance does. You definitely cannot grill a steak with a bread toaster, just like you can’t run an excellent, ultra-high performance storage services to serve the demanding AI and HPC applications on a general server platform. You have to have a storage appliance solution for High-Speed Storage.
That little Network Appliance® toaster award given out to exemplary employees stood vividly in my mind. The NetApp® tagline back then was “Fast, Simple, Reliable”. That solidifies my mindset for the high-speed storage in AI and HPC projects in present times.
DDN AI400X2 Turbo Appliance
Costs Benefits and Risks
I like to think about what the end users are thinking about. There are investments costs involved, and along with it, risks to the investments as well as their benefits. Let’s just simplify and lump them into Cost-Benefits-Risk analysis triangle. These variables come into play in the decision making of AI and HPC projects.
NAS (network attached storage) is obviously the file-level workhorse for shared resources in the network of any organization.SMB (server message block) for Windows environments and NFS (network file system) for Linux platforms are the 2 most prominent protocols that rule the NAS world. Of course we have SMB implementations in the form of Samba and others in non-Windows, Linux and NFS implementations in Windows as well.
As the versions of both network file sharing protocols iterated, present versions of SMB v3.x and NFS v4.x (NFS v3 on the supported Linux kernel version) on the client-side have evolved well. Both now have enhanced resiliency and performance improvement features. And there is an underlying similarity of both implementations. This blog looks at the client-side architectures of both.
One TCP connection
NAS is a client-server architecture. Over the network, NAS clients (SMB or NFS) access their corresponding NAS server(s) – SMB or NFS server(s) – through the TCP/IP network.
NAS client-server architecture (Credit: https://hypertecsp.com/en-CA/knowledge-base/nas-vs-san/)
One very important key starting point to note is the use of one TCP connection per NAS client to the NAS server relationship. For both SMB and NFS, there is just one TCP link between client and the server even if there are several SMB mapped shares or NFS mount points respectively on the clients.
For a long time, this one TCP connection is sufficient for the NAS traffic. But as the network file accesses grow, this connection becomes both a single point of failure as well as a performance bottleneck.
I am on a learning streak again. The most prominent technology that keeps landing on my tray at present is, of course, Artificial Intelligence (AI).
AI is hot. Very hot. And overhyped. Everyone is an expert nowadays. Yeah, right. Not me.
Underneath that glossy veneer of the AI hype, there are much going on behind the scenes to make AI great. The 2 areas I have been involved in and practiced for a long time are data infrastructure a.k.a. storage, and data management. And both are playing prominent parts in the advancement of the AI ecosystem. This makes me very excited.
I am no expert, but learning from various sources is already telling me that AI is pushing both storage and data management harder than ever before, much harder than traditional enterprise on-premises use cases and even the cloud computing applications. I ask myself, “where do I start my learning again?” as I journal my process.
Storage performance in a Data Pipeline
Speed of how AI responds is Trust. The faster it is to the accurate and relevant responses will build trust in AI. To get to the speed that we want is not an easy thing, and storage a.k.a. data infrastructure is doing its part. I pick up my learning from understanding the AI pipeline. One early help comes from my friend, Gina Rosenthal, who attended the Solidigm‘s presentation at AI Field Day in February 2024. Her article, titled “Why storage matters for AI – Solidigm“, kickstarted my learning juices again.
I was particularly captured by this slide in Gina’s article. It defines the laborious path data takes to become useful for AI applications.
Storage and the 5 stages of AI Work (reference: Solidigm presentation on AI Field Day)
I was listening to several storage luminaries in the GestaltIT’s podcast “No one understands Storage anymore” a few of weeks ago. Around the minute of 11.09 in the podcast, Dr. J. Metz, SNIA® Chair, brought up this is powerful quote “Storage does not mean Capacity“. It struck me, not in a funny way. It is what it is, and it something I wanted to say to many who do not understand the storage solutions they are purchasing. It exemplifies what is wrong in the many organizations today in their understanding of investing in a storage infrastructure project.
This is my pet peeve. The first words uttered in most, if not all storage requirements in my line of work are, “I want this many Terabytes of storage“. There are no other details and context of what the other requirement factors are, such as availability, performance, future growth, etc. Or even the goals to achieve when purchasing a storage system and operating it. What is the improvement they are looking for?What are the problems to solve?
Where is the OKR?
It pains me to say this. For the folks who have in the IT industry for years, both end users and IT purveyors alike, most are absolutely clueless about OKR (Objectives and Key Results) for their storage infrastructure project. Many cannot frame the data challenges they are facing, and they have no idea where to go next. There is no alignment. There is no strategy. Even worse, there is no concept of how their storage infrastructure investments will improve their business and operations.
Just the other day, one company director from a renown IT integrator here in Malaysia came calling. He has been in the IT industry since 1989 (I checked his Linkedin profile), asking to for a 100TB storage quote. I asked a few questions about availability, performance, scalability; the usual questions a regular IT guy would ask. He has no idea, and instead of telling me he didn’t know, he gave me a runaround of this and that. Plenty of yada, yada nonsense.
In the end, I told him to buy a consumer grade storage appliance from Taiwan. I will just let him make a fool of himself in front of his customer since he didn’t want to take accountability of ensuring his customer get a proper enterprise storage solution in good faith. His customer is probably in the same mould as well.
Defensive Strategies as Data Foundations
A strong storage infrastructure foundation is vital for good Data Credibility. If you do the right things for your data, there is Data Value, and it will serve your business well. Both Data Credibility and Data Value create confidence. And Confidence equates Trust.
In order to create the defensive strategies let’s look at storage Availability, Protection, Accessibility, Management Security and Compliance. These are 6 of the 8 data points of the A.P.P.A.R.M.S.C. framework.
Offensive Strategies as Competitive Advantage
Once we have achieved stability of the storage infrastructure foundation, then we can turn over and drive towards storage Performance, Recovery, plus things like Scalability and Agility.
With a strong data infrastructure foundation, the organization can embark on the offensive, and begin their business transformation journey, knowing that their data is well run, protection, and performs.
Alignment with Data and Business Goals
Why are the defensive and offensive strategies requiring alignment to business goals?
The fact is simple. It is about improving the business and operations, and setting OKRs is key to measure the ROI (return of investment) of getting the storage systems and the solutions in place. It is about switching the cost-fearing (negative) mindset to a profit-conviction (positive) mindset.
For example, maybe the availability of the data to the business is poor. Maybe there is the need to have access to the data 24×7, because the business is going online. The simple measurable fact is we can move availability from 95% uptime to 99.99% uptime with an HA storage system.
Perhaps there are concerns about recoverability in the deluge of ransomware threats. Setting new RPO goals from 24 hours to 4 hours is a measurable objective to enhance data resiliency.
Or getting the storage systems to deliver higher performance from 350 IOPS to 5000 IOPS for the database.
What I am saying here is these data points are measurable, and they can serve as checkpoints for business and operational improvements. From a management perspective, these can be used as KPI (key performance index) to define continuous improvement of Data Confidence.
Furthermore, it is easy when a OKR dashboard is used to map the improvement markers when organizations use storage to move from point A to point B, where B equates to a new success milestone. The alignment sets the paths to the business targets.
Storage does not mean only Capacity
The sad part is what the OKRs and the measured goals alignments are glaringly missing in the minds of many organizations purchasing a storage infrastructure and data management solution. The people tasked to source a storage technology solution are not placing a set of goals and objectives. Capacity appears to be the only thing on their mind.
I am about to meet a procurement officer of a customer soon. She asked me this question “Why is your storage so expensive?” over email. I want to change her mindset, just like the many officers and C-levels who hold the purse strings.
Let’s frame the use storage infrastructure in the real world. Nobody buys a storage system just to keep data in there much like a puddle keeps stagnant water. Sooner or later the value of the data in the storage evaporates or the value becomes dull if the data is not used well in any ways, shape or form.
Storage systems and the interconnected pathways from on premises, to the next premises, to the edge and to the clouds serve the greater good for Data. Data is used, shared, shaped, improved, enhanced, protected, moved, and more to deliver Value to the Business.
Storage capacity is just one of the few factors to consider when investing in a storage infrastructure solution. In fact, capacity is probably the least important piece when considering a storage solution to achieve the company’s OKRs. If we think about it deeper, setting the foundation for Data in the defensive manner will help elevate value of the data to be promoted with the offensive strategies to gain the competitive advantage.
Storage infrastructure and storage solutions along with data management platforms may appear to be a cost to the annual budgets. If you know set the OKRs, define A to get to B, alignment the goals, storage infrastructure and the data management platforms and practices are investments that are worth their weight in gold. That is my guarantee.
On the flip side, ignoring and avoiding OKRs, and set the strategies without prudence will yield its comeuppance. Technical debts will prevail.
One of the best features I have been waiting for is dRAID or distributed RAID. I have written about it dRAID a couple of years back. It was announced in 2021, in OpenZFS 2.1, but we have not seen an commercial implementation of dRAID … until iXsystems™ TrueNAS® SCALE 23.10. Why am I so excited?
I have followed the technology since Isaac Huang presented dRAID at the OpenZFS Summit in 2015. Through the years ahead, I have seen Isaac presenting dRAID at the summits, and with each iteration, dRAID got closer and closer to be developed into OpenZFS. It was not until 2021, in OpenZFS 2.1 when dRAID became part of filesystem. And now, dRAID is finally in the TrueNAS® SCALE offering.
Knowing RAID resilvering
RAID rebuilding or reconstruction is a painful and potentially risky process. In OpenZFS and ZFS speak, this process is called resilvering. In simple laymen terms, when a drive (or drives) failed in a parity-based RAID volume (eg. RAID-Z1 or RAID-Z2 vdev), the data which was previously in the failed drive is recreated in the newly integrated spare drive. The structural integrity of the RAID volume (and the storage pool) is preserved but the data that was lost is painstakingly remade through the mathematical algorithm of the parity function of the RAID volume.
When hard disk drives were small in capacity like 2TB or less, the RAID resilvering process was probably faster to complete, returning the parity RAID volume to a normal, online state. But today, drives are 22TB and higher, leaving the traditional RAID resilvering process to take days and even weeks. This leads the RAID volume vulnerable to another possible drive failure, weakening the integrity of the RAID volume. Even worse, most of modern day storage arrays have many disk drives, into the thousands even. And yes, solid state drives would probably be faster in the resilvering, but the same mechanics pretty much apply in OpenZFS.
At the same time, the spare drives are assigned physically and designated to the OpenZFS storage pool, and are not part of the vdev until the resilvering process kicks in.
Yes, this is pretty much a physical process that takes time, computing resources and patience. Note the operative word of “physical” here.
dRAID resilvering
dRAID speeds up the RAID resilvering process several folds, returning the RAID volume (or vdev) much faster than traditional OpenZFS RAID resilvering process. It uses a logical (as opposed to physical) RAID layout concept and uses “logical spare drives”. Thus, there will be many spares “blocks” distributed across the entire dRAID zpool, as shown in the diagram below.
For the uninformed, Clustered SMB brings highly available SMB file sharing services to mission critical environments. More importantly, Clustered SMB is high availability in a scale-out clustered architecture.
My view beyond HA SMB
I am not familiar with Clustered SMB in a NAS (Network Attached Storage). The world I am more familiar with is either having CIFS/SMB file services on a dual controller storage appliance or running Windows File Sharing on an Microsoft® Clustered Service (MSCS). Typically in these 2 types of HA SMB services, the scale up architecture require a shared access to a consolidated storage volume. Behind the scenes, there are many mechanisms at play to ensure that one, and only one, storage controller or HA host can have write access capabilities at one time. The most common mechanism is the SCSI-3 Persistent Reservation or sometimes known as SCSI fencing, using the SPC-3 (SCSI Primary Command) primitives. The whole objective is to prevent 2 nodes or hosts to writing to the shared storage volume at the same time and other issues like split-brain.
Data Fabric was a big buzzword going back several years. I wrote a piece talking about Data Fabric, mostly NetApp®’s, almost 7 years ago, which I titled “The Transcendence of Data Fabric“. Regardless of storage brands and technology platforms, and each has its own version and interpretations, one thing holds true. There must be a one layer of Data Singularity. But this is easier said than done.
Fast forward to present. The latest buzzword is Edge-to-Core-Cloud or Cloud-to-Core-Edge. The proliferation of Cloud Computing services, has spawned beyond to multiclouds, superclouds and of course, to Edge Computing. Data is reaching to so many premises everywhere, and like water,data has found its way.
Edge-to-Core-to-Cloud (Gratitude thanks to https://www.techtalkthai.com/dell-technologies-opens-iot-solutions-division-and-introduces-distributed-core-architecture/)
The question on my mind is can we have a single storage platform to serve the Edge-to-Core-to-Cloud paradigm? Is there a storage technology which can be the seamless singularity of data? 7+ years onwards since my Data Fabric blog, The answer is obvious. Object Storage.
The ubiquitous object storage and the S3 access protocol
For a storage technology that was initially labeled “cheap and deep”, object storage has become immensely popular with developers, cloud storage providers and is fast becoming storage repositories for data connectors. I wrote a piece called “All the Sources and Sinks going to Object Storage” over a month back, which aptly articulate how far this technology has come.
But unknown to many (Google NASD and little is found), object storage started its presence in SNIA (it was developed in Carnegie-Mellon University prior to that) in the early 90s, then known as NASD (network attached secure disk). As it is made its way into the ANSI T10INCITS standards development, it became known as Object-based Storage Device or OSD.
The introduction of object storage services 16+ years ago by Amazon Web Services (AWS) via their Simple Storage Services (S3) further strengthened the march of object storage, solidified its status as a top tier storage platform. It was to AWS’ genius to put the REST API over HTTP/HTTPS with its game changing approach to use CRUD (create, retrieve, update, delete) operations to work with object storage. Hence the S3 protocol, which has become the de facto access protocol to object storage.
Yes, I wrote those 2 blogs 11 and 9 years ago respectively because I saw that object storage technology was a natural fit to the burgeoning new world of storage computing. It has since come true many times over.
I just got home from the wonderful iXsystems™ Sales Summit in Knoxville, Tennessee. The key highlight was to christian the opening of iXsystems™ Maryville facility, the key operations center that will house iX engineering, support and part of marketing as well. News of this can be found here.
iX datacenter in the new Maryville facility
Western Digital® has always been a big advocate of iX, and at the Summit, they shared their hard disk drives HDD, solid state drives SSD, and other storage platforms roadmaps. I felt like a kid a candy store because I love all these excitements in the disk drive industry. Who says HDDs are going to be usurped by SSDs?
Several other disk drive manufacturers, including Western Digital®, have announced larger capacity drives. Here are some news of each vendor in recent months
I haven’t had hands-on with Gluster for over a decade. My last blog about Gluster was in 2011, right after I did a proof-of-concept for the now defunct, Jaring, Malaysia’s first ISP (Internet Service Provider). But I followed Gluster’s development on and off, until I found out that Gluster was a feature in then upcoming TrueNAS® SCALE. That was almost 2 years ago, just before I accepted to offer to join iXsystems™, my present employer.
The eagerness to test drive Gluster (again) on TrueNAS® SCALE has always been there but I waited for SCALE to become GA. GA finally came on February 22, 2022. My plans for the test rig was laid out, and in the past few weeks, I have been diligently re-learning and putting up the scope to built a 6-node Gluster clustered storage with TrueNAS® SCALE VMs on Virtualbox®.
Gluster on OpenZFS with TrueNAS SCALE
Before we continue, I must warn that this is not pretty. I have limited computing resources in my homelab, but Gluster worked beautifully once I ironed out the inefficiencies. Secondly, this is not a performance test as well, for obvious reasons. So, this is the annals along with the trials and tribulations of my 6-node Gluster cluster test rig on TrueNAS® SCALE.
