Disaggregation and Composability vital for AI/DL models to scale

New generations of applications and workloads like AI/DL (Artificial Intelligence/Deep Learning), and HPC (High Performance Computing) are breaking the seams of entrenched storage infrastructure models and frameworks. We cannot continue to scale-up or scale-out the storage infrastructure to meet these inundating fluctuating I/O demands. It is time to look at another storage architecture type of infrastructure technology – Composable Infrastructure Architecture.

Infrastructure is changing. The previous staid infrastructure architecture parts of compute, network and storage have long been thrown of the window, precipitated by the rise of x86 server virtualization almost 20 years now. It triggered a tsunami of virtualizing everything, including storage virtualization, which eventually found a more current nomenclature – Software Defined Storage. Both storage virtualization and software defined storage (SDS) are similar and yet different and should be revered through different contexts and similar goals. This Tech Target article laid out both nicely.

As virtualization raged on, converged infrastructure (CI) which evolved into hyperconverged infrastructure (HCI) went fever pitch for a while. Companies like Maxta, Pivot3, Atlantis, are pretty much gone, with HPE® Simplivity and Cisco® Hyperflex occasionally blipped in my radar. In a market that matured very fast, HCI is now dominated by Nutanix™ and VMware®, with smaller Microsoft®, Dell EMC® following them.

From HCI, the attention of virtualization has shifted something more granular, more scalable in containerization. Despite a degree of complexity, containerization is taking agility and scalability to the next level. Kubernetes, Dockers are now mainstay nomenclature of infrastructure engineers and DevOps. So what is driving composable infrastructure? Have we reached the end of virtualization? Not really.

Evolution of infrastructure. Source: IDC

It is just that one part of the infrastructure landscape is changing. This new generation of AI/ML workloads are flipping the coin to the other side of virtualization. As we see the diagram above, IDC brought this mindset change to get us to Think Composability, the next phase of Infrastructure.

Continue reading

Stating the case for a Storage Appliance approach

I was in Indonesia last week to meet with iXsystems™‘ partner PT Maha Data Solusi. I had the wonderful opportunity to meet with many people there and one interesting and often-replayed question arose. Why aren’t iX doing software-defined-storage (SDS)? It was a very obvious and deliberate question.

After all, iX is already providing the free use of the open source TrueNAS® CORE software that runs on many x86 systems as an SDS solution and yet commercially, iX sell the TrueNAS® storage appliances.

This argument between a storage appliance model and a storage storage only model has been debated for more than a decade, and it does come into my conversations on and off. I finally want to address this here, with my own views and opinions. And I want to inform that I am open to both models, because as a storage consultant, both have their pros and cons, advantages and disadvantages. Up front I gravitate to the storage appliance model, and here’s why.

My story of the storage appliance begins …

Back in the 90s, most of my work was on Fibre Channel and NFS. iSCSI has not existed yet (iSCSI was ratified in 2003). It was almost exclusively on the Sun Microsystems® enterprise storage with Sun’s software resell of the Veritas® software suite that included the Sun Volume Manager (VxVM), Veritas® Filesystem (VxFS), Veritas® Replication (VxVR) and Veritas® Cluster Server (VCS). I didn’t do much Veritas® NetBackup (NBU) although I was trained at Veritas® in Boston in July 1997 (I remembered that 2 weeks’ trip fondly). It was just over 2 months after Veritas® acquired OpenVision. Backup Plus was the NetBackup.

Between 1998-1999, I spent a lot of time working Sun NFS servers. The prevalent networking speed at that time was 100Mbits/sec. And I remember having this argument with a Sun partner engineer by the name of Wong Teck Seng. Teck Seng was an inquisitive fella (still is) and he was raving about this purpose-built NFS server he knew about and he shared his experience with me. I detracted him, brushing aside his always-on tech orgasm, and did not find great things about a NAS storage appliance. Auspex™ was big then, and I knew of them.

I joined NetApp® as Malaysia’s employee #2. It was an odd few months working with a storage appliance but after a couple of months, I started to understand and appreciate the philosophy. The storage Appliance Model made sense to me, even through these days.

Continue reading

Computational Storage embodies Data Velocity and Locality

I have been earnestly observing the growth of Computational Storage for a number of years now.  It was known by several previous names, with the name “in-situ data processing” stuck with me the most. The Computational Storage nomenclature became more cohesive when SNIA® put together the CMSI (Compute Memory Storage Initiative) some time back. This initiative is where several standards bodies, the major technology players and several SIGs (special interest groups) in SNIA® collaborated to advance Computational Storage segment in the storage technology industry we know of today.

The use cases for Computational Storage are burgeoning, and the functional implementations of Computational Storage are becoming vital to tackle the explosive data tsunami. In 2018 IDC, in its Worldwide Global Datasphere Forecast 2021-2025 report, predicted that the world will have 175 ZB (zettabytes) of data. That number, according to hearsay, has been revised to a heady figure of 250ZB, given the superlative rate data is being originated, spawned and more.

Computational Storage driving factors

If we take the Computer Science definition of in-situ processing, Computational Storage can be distilled as processing data where it resides. In a nutshell, “Bring Compute closer to Storage“. This means that there is a processing unit within the storage subsystem which does not require the host CPU to perform processing. In a very simplistic manner, a RAID card in a storage array can be considered a Computational Storage device because it performs the RAID functions instead of the host CPU. But this new generation of Computational Storage has much more prowess than just the RAID function in a RAID card.

There are many factors in Computational Storage that make a lot sense. Here are a few:

  1. Voluminous data inundate the centralized architecture of the cloud platforms and the enterprise systems today. Much of the data come from end point devices – mobile devices, sensors, IoT, point-of-sales, video cameras, et.al. Pre-processing the data at the origin data points can help filter the data, reduce the size to be processed centrally, and secure the data before they are ingested into the central data processing systems
  2. Real-time processing of the data at the moment the data is received gives the opportunity to create the Velocity of Data Analytics. Much of the data do not need to move to a central data processing system for analysis. Often in use cases like autonomous vehicles, fraud detection, recommendation systems, disaster alerts etc require near instantaneous responses. Performing early data analytics at the data origin point has tremendous advantages.
  3. Moore’s Law is waning. The CPU (central processing unit) is no longer the center of the universe. We are beginning to see CPU offloading technologies to augment the CPU’s duties such as compression, encryption, transcoding and more. SmartNICs, DPUs (data processing units), VPUs (visual processing units), GPUs (graphics processing units), etc have come forth to formulate a new computing paradigm.
  4. Freeing up central resources with Computational Storage also accelerates the overall distributed data processing in the whole data architecture. The CPU and the adjoining memory subsystem are less required to perform context switching caused by I/O interrupts as in most of the compute/storage architecture today. The total effect relieves the CPU and giving back more CPU cycles to perform higher processing tasks, resulting in faster performance overall.
  5. The rise of memory interconnects is enabling a more distributed computing fabric of data processing subsystems. The rising CXL (Compute Express Link™) interconnect protocol, especially after the Gen-Z annex, has emerged a force to be reckoned with. This rise of memory interconnects will likely strengthen the testimony of Computational Storage in the fast approaching future.

Computational Storage Deployment Models

SNIA Computational Storage Universe in 2019

Continue reading

How well do you know your data and the storage platform that processes the data

Last week was consumed by many conversations on this topic. I was quite jaded, really. Unfortunately many still take a very simplistic view of all the storage technology, or should I say over-marketing of the storage technology. So much so that the end users make incredible assumptions of the benefits of a storage array or software defined storage platform or even cloud storage. And too often caveats of turning on a feature and tuning a configuration to the max are discarded or neglected. Regards for good storage and data management best practices? What’s that?

I share some of my thoughts handling conversations like these and try to set the right expectations rather than overhype a feature or a function in the data storage services.

Complex data networks and the storage services that serve it

I/O Characteristics

Applications and workloads (A&W) read and write from the data storage services platforms. These could be local DAS (direct access storage), network storage arrays in SAN and NAS, and now objects, or from cloud storage services. Regardless of structured or unstructured data, different A&Ws have different behavioural I/O patterns in accessing data from storage. Therefore storage has to be configured at best to match these patterns, so that it can perform optimally for these A&Ws. Without going into deep details, here are a few to think about:

  • Random and Sequential patterns
  • Block sizes of these A&Ws ranging from typically 4K to 1024K.
  • Causal effects of synchronous and asynchronous I/Os to and from the storage

Continue reading

The burgeoning world of NVMe

When I wrote this article “Let’s smoke this storage peace pipe” 5 years ago, I quoted:

NVMe® and NVM®eF‰, as it evolves, can become the Great Peacemaker and bringing both divides and uniting them into a single storage fabric.

I envisioned NVMe® and NVMe®oF™ setting the equilibrium at the storage architecture level, finishing the great storage fabric into one. This balance in the storage ecosystem at the storage interface specifications and language-protocol level has rapidly unifying storage today, and we are already seeing the end-to-end NVMe paths directly from the PCIe bus of one host to another, via networks over Ethernet (with RoCE, iWARP, and TCP flavours) and Fibre Channel™. Technically we can have an end point device, example a tablet, talking the same NVMe language to its embedded storage as well as a cloud NVMe storage in an exascale storage far, far away. In the past, there were just too many bridges, links, viaducts, aqueducts, bypasses, tunnels, flyovers to cross just to deliver a storage command, or a data in a formats, encased and encoded (and decoded) in so many different ways.

Colours in equilibrium, like the rainbow

Simple basics of NVMe®

SATA (Serial Attached ATA) and SAS (Serial Attached SCSI) are not optimized for solid state devices. besides legacy stuff like AHCI (Advanced Host Controller Interface) in SATA, and archaic SCSI-3 primitives in SAS, NVM® has so much to offer. It can achieve very high bandwidth and support 65,535 I/O queues, each with a queue depth of 65,535. The queue depth alone is a massive jump compared to SAS which has a queue depth limit of 256.

A big part of this is how NVMe® handles I/O processing. It has a submission queue (SQ) and a completion queue (CQ), and together they are know as a Queue Pair (QP). The NVMe® controller handles tens of thousands at I/Os (reads and writes) simultaneously, alerted to switch between each SQ and CQ very quickly using the MSI or MSI-X interrupt. Think of MSI and MSI-X as a service bell, a hardware register that informs the NVM® controller when there are requests in the SQ, and informs the hosts that there are completed requests in the CQ. There will be plenty of “dings” by the MSI-X service register but the NVMe® controller can perform it very well, with some smart interrupt coalescing.

NVMe I/O processing

NVMe® 1.1, as I recalled, used to be have 3 admin commands and 10 base commands, which made it very lightweight compared to SCSI-3. However, newer commands were added to NVMe® 2.0 specifications included command sets fo key-value operations and zoned named space.

Continue reading

The future of Fibre Channel in the Cloud Era

The world has pretty much settled that hybrid cloud is the way to go for IT infrastructure services today. Straddled between the enterprise data center and the infrastructure-as-a-service in public cloud offerings, hybrid clouds define the storage ecosystems and architecture of choice.

A recent Blocks & Files article, “Broadcom server-storage connectivity sales down but recovery coming” caught my attention. One segment mentioned that the server-storage connectivity sales was down 9% leading me to think “Is this a blip or is it a signal that Fibre Channel, the venerable SAN (storage area network) protocol is on the wane?

Fibre Channel Sign

Thus, I am pondering the position of Fibre Channel SANs in the cloud era. Where does it stand now and in the near future? Continue reading

Storage IO straight to GPU

The parallel processing power of the GPU (Graphics Processing Unit) cannot be denied. One year ago, nVidia® overtook Intel® in market capitalization. And today, they have doubled their market cap lead over Intel®,  [as of July 2, 2021] USD$510.53 billion vs USD$229.19 billion.

Thus it is not surprising that storage architectures are changing from the CPU-centric paradigm to take advantage of the burgeoning prowess of the GPU. And 2 announcements in the storage news in recent weeks have caught my attention – Windows 11 DirectStorage API and nVidia® Magnum IO GPUDirect® Storage.

nVidia GPU

Exciting the gamers

The Windows DirectStorage API feature is only available in Windows 11. It was announced as part of the Xbox® Velocity Architecture last year to take advantage of the high I/O capability of modern day NVMe SSDs. DirectStorage-enabled applications and games have several technologies such as D3D Direct3D decompression/compression algorithm designed for the GPU, and SFS Sampler Feedback Streaming that uses the previous rendered frame results to decide which higher resolution texture frames to be loaded into memory of the GPU and rendered for the real-time gaming experience.

Continue reading

Storage Performance Considerations for AI Data Paths

The hype of Deep Learning (DL), Machine Learning (ML) and Artificial Intelligence (AI) has reached an unprecedented frenzy. Every infrastructure vendor from servers, to networking, to storage has a word to say or play about DL/ML/AI. This prompted me to explore this hyped ecosystem from a storage perspective, notably from a storage performance requirement point-of-view.

One question on my mind

There are plenty of questions on my mind. One stood out and that is related to storage performance requirements.

Reading and learning from one storage technology vendor to another, the context of everyone’s play against their competitors seems to be  “They are archaic, they are legacy. Our architecture is built from ground up, modern, NVMe-enabled“. And there are more juxtaposing, but you get the picture – “We are better, no doubt“.

Are the data patterns and behaviours of AI different? How do they affect the storage design as the data moves through the workflow, the data paths and the lifecycle of the AI ecosystem?

Continue reading

Scaling new HPC with Composable Architecture

[Disclosure: I was invited by Dell Technologies as a delegate to their Dell Technologies World 2019 Conference from Apr 29-May 1, 2019 in the Las Vegas USA. Tech Field Day Extra was an included activity as part of the Dell Technologies World. My expenses, travel, accommodation and conference fees were covered by Dell Technologies, the organizer and I was not obligated to blog or promote their technologies presented at this event. The content of this blog is of my own opinions and views]

Deep Learning, Neural Networks, Machine Learning and subsequently Artificial Intelligence (AI) are the new generation of applications and workloads to the commercial HPC systems. Different from the traditional, more scientific and engineering HPC workloads, I have written about the new dawn of supercomputing and the attractive posture of commercial HPC.

Don’t be idle

From the business perspective, the investment of HPC systems is high most of the time, and justifying it to the executives and the investors is not easy. Therefore, it is critical to keep feeding the HPC systems and significantly minimize the idle times for compute, GPUs, network and storage.

However, almost all HPC systems today are inflexible. Once assigned to a project, the resources pretty much stay with the project, even when the workload processing of the project is idle and waiting. Of course, we have to bear in mind that not all resources are fully abstracted, virtualized and software-defined whereby you can carve out pieces of the hardware and deliver a percentage of that resource. Case in point is the CPU, where you cannot assign certain clock cycles of CPU to one project and another half to the other. The technology isn’t there yet. Certain resources like GPU is going down the path of Virtual GPU, and into the realm of resource disaggregation. Eventually, all resources of the HPC systems – CPU, memory, FPGA, GPU, PCIe channels, NVMe paths, IOPS, bandwidth, burst buffers etc – should be disaggregated and pooled for disparate applications and workloads based on demands of usage, time and performance.

Hence we are beginning to see the disaggregated HPC systems resources composed and built up the meet the diverse mix and needs of HPC applications and workloads. This is even more acute when a AI project might grow cold, but the training of AL/ML/DL workloads continues to stay hot

Liqid the early leader in Composable Architecture

Continue reading