Category Archives: Infiniband

Rethinking Storage OKRs for AI Data Infrastructure – Part 1

By | | , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , ,
Leave a comment

[ Preamble: This analysis focuses on my own journey as I incorporate my past experiences into this new market segment called AI Data Infrastructure, and gaining new ones.

There are many elements of HPC (High Performance Computing) at play here. Even though things such as speeds and feeds, features and functions crowd many conversations, as many enterprise storage vendors like to do, these conversations, in my opinion, are secondary. There are more vital and important operational technology and technical elements that an organization has to consider prudently, vis-a-vis to ROIs (returns of investments). They involve asking the hard questions beyond the marketing hype and fluff. I call these elements of consideration Storage Objectives and Key Results (OKRs) for AI Data Infrastructure.

I had to break this blog into 2 parts. It has become TL;DR-ish. This is Part 1 ]

I have just passed my 6-month anniversary with DDN. Coming into the High Performance Storage System (HPSS) market segment, with the strong focus on the distributed parallel filesystem of Lustre®, there was a high learning curve for me. I spend over 3 decades in Enterprise Storage, with some of the highest level of storage technologies there were in that market segment. And I have already developed my own approach to enterprise storage, based on the A.P.P.A.R.M.S.C.. That was already developed and honed from 25 years ago.

The rapid adoption of AI has created a technology paradigm shift. Artificial Intelligence (AI) came in and blurred many lines. It also has been evolving my thinking when it comes to storage for AI. There is also a paradigm shift in my thoughts, opinions and experiences as well.

AI has brought HPSS technologies like Lustre® in DDN EXAscaler platform , proven in the Supercomputing world, to a new realm – the AI Data Infrastructure market segment. On the other side, many enterprise storage vendors aspire to be a supplier to the AI Data Infrastructure opportunities as well. This convergence from the top storage performers for Supercomputing, in the likes of DDN, IBM® (through Storage Scale), HPE® (through Cray, which by-the-way often uses the open-source Lustre® edition in its storage portfolio), from the software-defined storage players in Weka IO, Vast Data, MinIO, and from the enterprise storage array vendors such as NetApp®, Pure Storage®, and Dell®.

[ Note that I take care not to name every storage vendor for AI because many either do OEMs or repacking and rebranding of SDS technology into their gear such as HPE® GreenLake for Files and Hitachi® IQ. You can Google to find out who the original vendors are for each respectively. There are others as well. ]

In these 3 simplified categories (HPSS, SDS, Enterprise Storage Array), I have begun to see a pattern of each calling its technology as an “AI Data Infrastructure”. At the same time, I am also developing a new set of storage conversations for the AI Data Infrastructure market segment, one that is based on OKRs (Objectives and Key Results) rather than just features, features and more features that many SDS and enterprise storage vendors like to tout. Here are a few thoughts that we should look for when end users are considering a high-speed storage solution for their AI journey.

AI Data Infrastructure

GPU is king

In the AI world, the GPU infrastructure is the deity at the altar. The utilization rate of the GPUs is kept at the highest to get the maximum compute infrastructure return-on-investment (ROI). Keeping the GPUs resolutely busy is a must. HPSS is very much part of that ecosystem.

These are a few OKRs I would consider the storage or data infrastructure for AI.

  • Reliability
  • Speed
  • Power Efficiency
  • Security

Let’s look at each one of them from the point of view of a storage practitioner like me.

Continue reading →

Accelerated Data Paths of High Performance Storage is the Cornerstone of building AI

By | | , , , , , , , , , , , , , , , , , , , , , , , , , , , , ,
Leave a comment

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.

Continue reading →

The All-Important Storage Appliance Mindset for HPC and AI projects

By | | , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , ,
Leave a comment

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.

Continue reading →

I built a 6-node Gluster cluster with TrueNAS SCALE

By | | , , , , , , , , , , , , , , , , , , , , , , ,
10 Comments

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.

Continue reading →

The burgeoning world of NVMe

By | | , , , , , , , , , , , , , , , , ,
1 Comment

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 →

Is Software Defined right for Storage?

By | | , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , ,
Leave a comment

George Herbert Leigh Mallory, mountaineer extraordinaire, was once asked “Why did you want to climb Mount Everest?“, in which he replied “Because it’s there“. That retort demonstrated the indomitable human spirit and probably exemplified best the relationship between the human being’s desire to conquer the physical limits of nature. The software of humanity versus the hardware of the planet Earth.

Juxtaposing, similarities can be said between software and hardware in computer systems, in storage technology per se. In it, there are a few schools of thoughts when it comes to delivering storage services with the notable ones being the storage appliance model and the software-defined storage model.

There are arguments, of course. Some are genuinely partisan but many a times, these arguments come in the form of the flavour of the moment. I have experienced in my past companies touting the storage appliance model very strongly in the beginning, and only to be switching to a “software company” chorus years after that. That was what I meant about the “flavour of the moment”.

Software Defined Storage

Continue reading →

Glusterific!

By | | , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , ,
Leave a comment

A conversation with a storage executive last week brought up Gluster, a clustered file system I have not explored in many years. I had one interaction months before its acquisition by RedHat® in 2011.

I remembered the Gluster demo at Jaring over a video call, because I was the lead consultant pitching the scale-out NAS solution. It did not go well, and there were “bugs” which made the Head of IT flinched in her seat. Despite Jaring being Malaysia’s technology trailblazer, the impression of Gluster was forgettable. I stayed on the GlusterFS architecture a little while and then it dropped off my radar.

Gluster Logo Scale Out NAS

Gluster Scale Out NAS

But after the conversation last week, I am elated to revive my interest in Gluster, knowing that something big and impressive in coming into the fore very soon. Studying the architecture (again!), there are 2 parts of Gluster which excite me. One is the Brick and the other is the lack of a Metadata service.

Continue reading →

Storage Performance Considerations for AI Data Paths

By | | , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , ,
1 Comment

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

By | | , , , , , , , , , , , , , , , , , , , , , , , , , , , , ,
2 Comments

[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 →

WekaIO controls their performance destiny

By | | , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , ,
3 Comments

[Preamble: I have been invited by GestaltIT as a delegate to their Tech Field Day for Storage Field Day 18 from Feb 27-Mar 1, 2019 in the Silicon Valley USA. My expenses, travel and accommodation were covered by GestaltIT, 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]

I was first introduced to WekaIO back in Storage Field Day 15. I did not blog about them back then, but I have followed their progress quite attentively throughout 2018. 2 Storage Field Days and a year later, they were back for Storage Field Day 18 with a new CTO, Andy Watson, and several performance benchmark records.

Blowout year

2018 was a blowout year for WekaIO. They have experienced over 400% growth, placed #1 in the Virtual Institute IO-500 10-node performance challenge, and also became #1 in the SPEC SFS 2014 performance and latency benchmark. (Note: This record was broken by NetApp a few days later but at a higher cost per client)

The Virtual Institute for I/O IO-500 10-node performance challenge was particularly interesting, because it pitted WekaIO against Oak Ridge National Lab (ORNL) Summit supercomputer, and WekaIO won. Details of the challenge were listed in Blocks and Files and WekaIO Matrix Filesystem became the fastest parallel file system in the world to date.

Control, control and control

I studied WekaIO’s architecture prior to this Field Day. And I spent quite a bit of time digesting and understanding their data paths, I/O paths and control paths, in particular, the diagram below:

Starting from the top right corner of the diagram, applications on the Linux client (running Weka Client software) and it presents to the Linux client as a POSIX-compliant file system. Through the network, the Linux client interacts with the WekaIO kernel-based VFS (virtual file system) driver which coordinates the Front End (grey box in upper right corner) to the Linux client. Other client-based protocols such as NFS, SMB, S3 and HDFS are also supported. The Front End then interacts with the NIC (which can be 10/100G Ethernet, Infiniband, and NVMeoF) through SR-IOV (single root IO virtualization), bypassing the Linux kernel for maximum throughput. This is with WekaIO’s own networking stack in user space. Continue reading →