104: GreyBeards talk new cloud defined (shared) storage with Siamak Nazari, CEO Nebulon

Ray has known Siamak Nazari (@NebulonInc), CEO Nebulon for three companies now but has rarely had a one (two) on one discussion with him. With Nebulon just emerging from stealth (a gutsy move during the pandemic), the GreyBeards felt it was a good time to get Siamak on the show to tell us what he’s been up to. Turns out he and Nebulon decided it was time to completely rethink/rearchitect shared storage for the new data center.

At his prior company, Siamak spent a lot of time with many customers discussing the problems they had dealing with the complexity of managing, provisioning and maintaining multiple shared storage arrays. Somewhere in all those discussions Siamak saw this as a problem that needed a radical solution. If we could just redo shared storage from the ground up, there might be a solution to all these problems.

Redefining shared storage

Nebulon’s new approach to shared storage starts with an SPU card which replaces SAS RAID cards in a server. But instead of creating SAS RAID groups, the SPU creates a shareable, enterprise class, pool of storage across a throng of servers.

They call a collection of servers with SPUs, Cloud Defined Storage (CDS) and it creates a Nebulon nPod. An nPod essentially consists of multiple servers with SPU cards, with or without attached SSD storage, that are provisioned, managed and monitored via the cloud. Nebulon nPod servers are elements or nodes of a shared storage pool across all interconnected SPU servers in a data center.

In an SPU server with local (SAS, SATA, NVMe) SSD storage, the SPU creates an erasure coded pool of storage which can be used to serve (SAS) LUNs to this or any other SPU attached server in the nPod. In a SPU server without local SSD storage, the SPU provides access to any other SPU server shared storage in the nPod. Nebulon nPods only works with flash storage, it doesn’t support spinning media.

The SPU can supply boot storage for its server. There’s no need to have the CPU running OS code to use nPod shared storage. Yes, the SPU needs power and an active PCIe bus to work, but the functionality of an SPU doesn’t require an operational OS to work. The SPU provides a SAS LUN interface to server CPUs.

Each SPU has dual port access to an inter-cluster (25GbE) interconnect that connects all SPUs to the nPod. The nPod inter-cluster protocol is proprietary but takes advantage of standard TCP/IP services across the network with standard 25GbE switching.

The SPU firmware insures that it stays connected as long as power is available to the server. Customers can have more than one SPU in a server but these would be used for more IO performance. Each SPU also has 32GB of NVRAM for caching purposes and it’s also used for power fail fault tolerance.

In the unlikely case that the server and SPU are completely down (e.g. power outage), clients can still access that SPUs data storage, if it was mirrored (see below). When the SPU server comes back up, it will be resynched with any data that had been changed.

Other Nebulon storage features

Nebulon supports data-at-rest encryption, compression and deduplication for customer data. That way customer data is never in plain text as it travels across the nPod or even within the server from the SPU to SSD storage. Also any customer data written to an nPod can be optionally mirrored and as noted above, is protected via erasure coding.

The SPU also supports snapshotting of customer LUN data. So clients can take copies of LUNs and use these for backups, test, dev, etc. SPUs also support asynchronous or synchronous replication between nPods. For synchronous replication and mirrored data, the originating host only sees the IO complete after the data has been received at the target SPU or nPod.

Metadata for the nPod that defines LUN configurations and which server has LUN data is kept across the cluster in each SPU. But metadata on the location of user data within a server is only kept in that server’s SPU.

We asked Siamak whether nPods support SCM (storage class memory). He said not yet, but they’re looking at SCM NVMe storage for use as a potential metadata and data cache for SPUs.

Nebulon Application Centric storage

All the above storage features are present in most enterprise class storage systems. But what sets Nebulon apart from all other shared storage arrays is that their control plane is entirely in the cloud. That is customers point their browser to Nebulon’s control plane and use it to configure, provision and manage the nPod storage pool. Nebulon supports application templates that can be used to configure nPod storage to support standardized applications, such as VMware VMs, MongoDB, persistent storage for K8S containers, bare metal Linux apps, etc.

With the nPod’s control plane in the cloud it makes provisioning, managing and monitoring storage services much more agile. Nebulon can literally roll out new control plane updatesy to their install base on an almost daily basis. Just like any other cloud based or SAAS application. Customers receive the updated nPod control plane functionality by simply refreshing their browser page.

Nebulon’s GoToMarket

Near the end of our podcast, we asked Siamak about how Nebulon was going to access the market. Nebulon’s goto market is to use server OEMs. That is, they have signed agreements with two (and working on a third) server vendors to sell SPU cards with Nebulon control plane access.

During server purchases, customers configure their servers but now along with SAS RAID card options they will now see an Nebulon SPU option. OEM server vendors will bundle SPU hardware and Nebulon control plane access along with all other server components such as CPU’s, SSDs, NICs, etc, This way, the customer will receive a pre-installed SPU card in their server and will be ready to configure nPod LUNs as soon as the server powers on in their network.

Nebulon will go GA in the 3rd quarter.

The podcast ran ~43 minutes. Siamak has always been a pleasure to talk with and is very knowledgeable about the problems customers have in today’s data center environments. Nebulon has given him and his team the way to rethink storage and address these serious issues. Matt and I had a good time talking with Siamak. Listen to the podcast to learn more.

This image has an empty alt attribute; its file name is Spotify_Logo_CMYK_Black-1024x307.png
This image has an empty alt attribute; its file name is play_prism_hlock_2x-300x64.png
This image has an empty alt attribute; its file name is Subscribe_on_iTunes_Badge_US-UK_110x40_0824.png

Siamak Nazari, CEO Nebulon

Siamak Nazari is the CEO and Co-founder of Nebulon. Siamak has over 25 years of experience working on distributed and highly available systems.

In his position as HPE Fellow and VP, he was responsible for setting technical direction for HPE 3PAR and its portfolio of software and hardware. He worked on HPE 3PAR technology from 2000 to 2018, responsible for designing and implementing distributed memory management and the high availability features of the system.

Prior to joining 3PAR, Siamak was the technical lead for distributed highly available Proxy Filesystem (pxfs) of Sun Cluster 3.0.

92: Ray talks AI with Mike McNamara, Sr. Manager, AI Solution Mkt., NetApp

Sponsored By: NetApp

NetApp’s been working in the AI DL (deep learning) space for a long time now and announced their partnership with NVIDIA DGX systems, back in August of 2018. At NetApp Insight, this week they were showing off their new NVIDIA DGX systems reference architectures. These architectures use NetApp AFF A800 storage (for more info on AI DL, checkout Ray’s Learning Machine (deep) Learning posts – part 1, – part 2 and – part3).

Besides the ONTAP AI systems, NetApp also offers

  • FlexPod AI solution based on their partnership with Cisco using UCS C480 ML M5 rack servers which include 8 NVIDA Tesla V100 GPUs and also features NetApp AFF A800 storage for use in core AI DL.
  • NetApp HCI has two configurations with 2- or 3-NVIDIA GPUs that come in 1U or 2U rack servers and run VMware vSphere or RedHad OpenStack/OpenShift software hypervisors suitable for edge or core AI DL.
  • E-series reference architecture that uses the BeeGFS parallel file system and offers InfiniBAND data access for HPC or core AI DL.

On the conference floor, NetApp showed AI DL demos for automotive, financial services, Public Sector and healthcare verticals. They also had a facial recognition application running that could estimate your age and emotional state (I didn’t try it, but Mike said they were hedging the model so it predicted a lower age).

Mike said one healthcare solution was focused on radiological image scans, to identify pathologies from x-Ray, MRI, or CAT scan images. Mike mentioned there was a lot of radiological technologists burn-out due to the volume of work caused by the medical imaging explosion over the last decade or so. Mike said image analysis is something that h AI DL can perform very effectively and doing so would improve the accuracy and reduce the volume of work being done by technologists.

He also mentioned another healthcare application that uses an AI DL app to count TB cells in blood samples and estimate the extent of TB infections. Historically, this has been time consuming, error prone and hard to do in the field. The app uses a microscope with a smart phone and can be deployed and run anywhere in the world.

Mike mentioned a genomics AI DL application that examined DNA sequences and tried to determine its functionality. He also mentioned a retail AI DL facial recognition application that would help women “see” what they would look like with different makeup on.

There was a lot of discussion on NetApp Cloud services at the show, such as Cloud Volume Services and Azure NetApp File (ANF). Both of these could easily be used to implement an AI DL application or be part of an edge to core to cloud data flow for an AI DL application deployment using NetApp Data Fabric.

NetApp also announced a new, all flash StorageGRID appliance that was targeted at heavy IO intensive uses of object store like AI DL model training and data analytics.

Finally, Mike mentioned NetApp’s ecosystem of partners working in the AI space to help customers deploy AI DL algorithms in their industries. Some of these include:

  1. Flexential, Try and Buy AI so that customers could bring them in to supply AI DL expertise to generate an AI DL application using customer data and deploy it on customer cloud or on prem infrastructure .
  2. Core Scientific, AI-as-a-Service, so that customers could purchase a service to implement an AI DL application using customer data and running on Core Scientific infrastructure..
  3. Scale Matrix, Mobile data center AI, so that customers could create an AI DL application and run it on Scale Matrix infrastructure that was transported to wherever the customer wanted it to be run.

We recorded the podcast on the show floor, in a glass booth, so there’s some background noise (sorry about that, but can’t be helped). The podcast is ~27 minutes. Mike is a long time friend and NetApp product expert, recently working in AI DL solutions at NetApp. When I saw Mike at Insight, I just had to ask him about what NetApp’s been doing in the AI DL space. Listen to the podcast to learn more.

This image has an empty alt attribute; its file name is Subscribe_on_iTunes_Badge_US-UK_110x40_0824.png
This image has an empty alt attribute; its file name is play_prism_hlock_2x-300x64.png

Mike McNamara, Senior Manager AI Solution Marketing, NetApp

With over 25 years of data management product and solution marketing experience, Mike’s background includes roles of increasing responsibility at NetApp (10+ years), Adaptec, EMC and Digital Equipment Corporation. 

In addition to his past role as marketing chairperson for the Fibre Channel Industry Association, he was a member of the Ethernet Technology Summit Conference Advisory Board, a member of the Ethernet Alliance, and a regular contributor to industry journals, and a frequent speaker at events.

83: GreyBeards talk NVMeoF/TCP with Muli Ben-Yehuda, Co-founder & CTO and Kam Eshghi, VP Strategy & Bus. Dev., Lightbits Labs

This is the first time we’ve talked with Muli Ben-Yehuda (@Muliby), Co-founder & CTO and Kam Eshghi (@KamEshghi), VP of Strategy & Business Development, Lightbits Labs. Keith and I first saw them at Dell Tech World 2019, in Vegas as they are a Dell Ventures funded organization. The company has 70 (mostly engineering) employees and is based in Israel, with offices in NY and the Valley as well as elsewhere around the world. Kam was previously with (Dell) EMC DSSD and Muli’s spent years as a Master Inventor with IBM Research.

[This was Keith Townsend’s (@CTOAdvisor & The CTO Advisor), first time as a GreyBeard co-host and we had a great time with him on the show.]

I would have to say it was a far ranging discussion but focused on their software defined, NVMeoF/TCP storage. As you may recall we talked with Solarflare Communications last year who were also working on a NVMeoF/TCP, only in their case it was an accelerator board. After the recording, Muli said the hardware accelerator they have is their own design.

Why NVMeoF/TCP?

Most NVMeoF today, that uses Ethernet, requires RoCE or iWARP compatible NICs and switches. Lightbits Labs has long been active in the NVMeoF/RoCE-iWARP market place. Early on they noticed that enterprise and cloud service providers were reluctant to adopt NVMeoF technology because of the need to change out all their networking equipment to use it. This is what brought about their focus on NVMeoF/TCP.

The advantage of NVMeoF/TCP is that it can be run on any Ethernet NIC and switch available today. From Muli’s perspective, NVMeoF/TCP is going to become the next SAN of choice for the data center. They were active, early on, in the standards committee to push for NVMeoF/TCP adoption.

How does it work?

Their software defined solution runs LightOS® storage software, a Linux based package, and uses off the shelf, server hardware with persistent storage (Optane DC PM/SSDs, NV DIMMs, V-NAND, etc.). They use persistent memory for a FAST write buffer and a place where they can “mold” the written data into something that can be better written to backend NVMe SSDs.

One surprise about Lightbits solution is that it offers a decent set of data services. These include erasure coding, thin provisioning, wire-speed inline compression, QoS and wide striping. It seems like any of these can be disabled by a customers want. But they only add very little overhead. I think Muli mentioned one Lightbits customer with encrypted data that disabled compression.

Lightbits also offers a global FTL (flash translation layer), which means they control SSD addressing which maps data to physical/raw NAND locations at the storage system level. If done well, a global FTL can help improve flash endurance and may offer better write performance (through increased parallelism).

Lightbits claim to inline, wire speed data compression is premised on the use of more current CPUs with high (>=28) core counts in a storage server. If the storage server has older CPUs (<28 cores), they suggest you install their LightField™ hardware accelerator add in card. LightField offers a number of hardware based, performance accelerations in addition to compression speedups.

LightOS requires no host (client) software. Muli’s a long time Linux kernel contributor and indicated that the only thing LightOS needs is a current Linux Kernel (5.0 or later) which has the NVMeoF/TCP driver software (and persistent memory). Lightbits believes that it’s only a matter of time until other OSs also implement NVMeoF/TCP drivers.

Lightbits business considerations

Long term, Lightbits sees a need for compute-storage disaggregation in hyper scalar and enterprise cloud environments. Early on it was relatively easy to replicate servers with DAS storage but as NVMe SSDs came out the expense to do this throughout their >>1000 server environment starts to become exorbitant. If they only had an easy way to disaggregate their storage from compute and still enjoy all the performance advantages of DAS NVMe SSDS. With LightOS they can do that.

Lightbits can be sold today through Dell, as a partner solution, which means that Dell can integrate, test and validate their servers with LightField accelerator card and deliver that package to your data center. I believe you still need to purchase and install their LightOS software yourself.

Lightbits charges for LightOS software on a per storage node basis, but they have different charges based on the maximum number of NVMe SSD slots available is in a server. There is no capacity charge. They also offer worldwide service and support for LightOS software and LightField hardware.

It’s all about performance

From a performance perspective, one Fortune 500 hyper-scalar benchmarked their storage solution against a DAS NVMe server and found it added about 30 µsec to the IO latency as compare to DAS NVMe SSDs. From their perspective, the added data services, better endurance, and disaggregated compute-storage environment provided by LightOS more than made up for the additional overhead.

Finally, I asked about whether multiple LightOS storage servers could be clustered together. Muli intervened, after stating some legal stuff, said they were working on the next generation LightOS and it will support clustered storage servers, local data replication as well as distributed (across storage servers) erasure coding.

The podcast is a long one and runs over ~47 minutes. There was a lot to talk about and Kam and Muli seem to know it all. It was interesting to hear the history of their pivot to TCP. They seem to have the right technology to address the market. Listen to the podcast to learn more.

Muli Ben-Yehuda, Co-founder and CTO, Lightbits Labs

Muli Ben-Yehuda is the CTO and Co-Founder of Lightbits Labs, where he leads technological developments.

Prior to founding Lightbits, he was chief scientist at Stratoscale and a researcher and Master Inventor at IBM Research.

He holds an M.Sc. in Computer Science (summa cum laude) from the Technion — Israel Institute of Technology and a B.A. (cum laude) from the Open University of Israel.

He is a long time Linux kernel contributor and his code and ideas are most likely included in an operating system or hypervisor running near you. He is also one of the authors of the NVMe/TCP standard and technology. 

Kam Eshghi, VP Strategy & Business Development, Lightbits Labs

Kam joined Lightbits Labs from Dell EMC and has over 20yrs of experience in strategic marketing and business development with startups and public companies.

Most recently as VP of strategic alliances at startup DSSD, Kam led business development with technology partners and developed DSSD’s partnership with EMC, leading to EMC’s acquisition of DSSD.

Previously as Sr. Director of Marketing & Business Development at IDT, Kam built their NVMe Controller business from scratch. Previous to that, Kam worked in data center storage, compute and networking markets at HP, Intel, and Crosslayer Networks. 

Kam is a U.C. Berkeley and MIT graduate with a BS and MS in Electrical Engineering and Computer Science and an MBA.

79: GreyBeards talk AI deep learning infrastructure with Frederic Van Haren, CTO & Founder, HighFens, Inc.

We’ve talked with Frederic before (see: Episode #33 on HPC storage) but since then, he has worked for an analyst firm and now he’s back on his own again, at HighFens. Given all the interest of late in AI, machine learning and deep learning, we thought it would be a great time to catch up and have him shed some light on deep learning and what it needs for IT infrastructure.

Frederic has worked for HPC / Big Data / AI / IoT solutions in the speech recognition industry, providing speech recognition services for some of the largest organizations in the world. As I understand it, the last speech recognition AI application he worked on implemented deep learning.

A brief history of AI

Frederic walked the Greybeards through the history of AI from the dawn of computing (1950s) until the recent emergence of deep learning (2010).

He explained that, early on one could implement a chess playing program, using hand coded rules based on a chess expert’s playing technique. Later when machine learning came out, one could use statistical analysis on multiple games and limited rule creation to teach a AI machine learning system how to play chess. With deep learning (DL), all you have to do now is to feed a DL model all the games you have and it learns how to play chess well all by itself. No rule making needed.

AI DL training and deployment infrastructure

Frederic described some of the infrastructure and data needs for various phases of an industrial scale, AI DL workflow.

Training deep learning models takes data and the more, the better. Gathering/saving large amounts of data used for DL training is a massive write workload and at the end of that process, hopefully you have PB of data to work with.

Selecting DL training data from all those PBs, involves a lot of mixed read and write IO. In the end, one has selected and extracted the data to use to train your DL models.

During DL training, IO needs are all about heavy data read throughput. But there’s more, in the later half of the talk, Frederic talked about the need to keep expensive GPU cores busy and that requires sophisticated caching or Tier 0 storage supporting low latency IO.

Ray’s been doing a lot of blogging and other work on AI machine and deep learning (e.g., see Learning machine learning – parts 1, 2, & 3) so it was great to hear from Frederic, a real practitioner of the art. Frederic (with some of Ray’s help) explained the deep learning training process. But it wasn’t detailed enough for Howard, so per Howard’s request, we went deeper into how it really works.

Once you have a DL model trained and working within specifications (e.g., prediction accuracy), Frederic said deploying DL models into production involves creating two separate clusters. One devoted to deep learning model inferencing, which takes in data from the world and performs inferencing (prediction, classification, interpretations, etc.) and the other uses that information for model adaption to fine tune DL models for specific instances.

Adaption and inferencing were both read and write IO workloads and the performance of this IO was dependent on a specific model’s use

Model adaption would personalize model predictions for each and every person, car, genotype, etc. This would be done periodically (based on SLAs, e.g. every 4 hrs). After that, a new, adapted model could be introduced into production, adapted for that specific person/car/genotype.

If the adaption applied more generally, that data and its human-machine validated/vetted prediction, classification, interpretation, etc. would be added back into the DL model training set to be used the next time a full model training pass was to be done. Frederic said AI DL model training is never done.

Sometime later, all this DL training, production and adaption data needs to be archived for long term access.

We then discussed the recent offerings from NVIDIA and major storage vendors that package up a solution for AI deep learning. It seems we are seeing another iteration of Converged Infrastructure, only this time for AI DL.

Finally, over the course of Ray’s AI DL education, he had come to the belief that AI deep learning could be applied by anyone. Frederic corrected Ray stating that AI deep learning should be applied by anyone.

The podcast runs ~44 minutes. Frederic’s been an old friend of Howard’s and Ray’s, since before the last podcast. He’s one of the few persons in the world that the GreyBeards know that has real world experience in deploying AI DL, at industrial scale. Frederic’s easy to talk with and very knowledgeable about the intersection of Ai DL and IT infrastructure. Howard and I had fun talking with him again on this episode. Listen to the podcast to learn more. .

Frederic Van Haren

Frederic Van Haren is the Chief Technology Officer @ HighFens. He has over 20 years of experience in high tech and is known for his insights in HPC, Big Data and AI from his hands-on experience leading research and development teams. He has provided technical leadership and strategic direction in the Telecom and Speech markets.

He spent more than a decade at Nuance Communications building large HPC and AI environments from the ground up and is frequently invited to speak at events to provide his vision on the HPC, AI, and storage markets. Frederic has also served as the president of a variety of technology user groups promoting the use of innovative technology.

As an engineer, he enjoys working directly with engineering teams from technology vendors and on challenging customer projects.

Frederic lives in Massachusetts,  USA but grew up in the northern part of Belgium where he received his Masters in Electrical Engineering, Electronics and Automation.

78: GreyBeards YE2018 IT industry wrap-up podcast

In this, our yearend industry wrap up episode, we discuss trends and technology impacting the IT industry in 2018 and what we can see ahead for 2019 and first up is NVMeoF

NVMeoF has matured

In the prior years, NVMeoF was coming from startups, but last year it’s major vendors like IBM FlashSystem, Dell EMC PowerMAX and NetApp AFF releasing new NVMeoF storage systems. Pure Storage was arguably earliest with their NVMeoF JBOF.

Dell EMC, IBM and NetApp were not far behind this curve and no doubt see it as an easy way to reduce response time without having to rip and replace enterprise fabric infrastructure.

In addition, NVMeoFstandards have finally started to stabilize. With the gang of startups, standards weren’t as much of an issue as they were more than willing to lead, ahead of standards. But major storage vendors prefer to follow behind standards committees.

As another example, VMware showed off an NVMeoF JBOF for vSAN. A JBoF like this improves vSAN storage efficiency for small clusters. Howard described how this works but with vSAN having direct access to shared storage, it can reduce data and server protection requirements for storage. Especially, when dealing with small clusters of servers becoming more popular these days to host application clusters.

The other thing about NVMeoF storage is that NVMe SSDs have also become very popular. We are seeing them come out in everyone’s servers and storage systems. Servers (and storage systems) hosting 24 NVMe SSDs is just not that unusual anymore. For the price of a PCIe switch, one can have blazingly fast, direct access to a TBs of NVMe SSD storage.

HCI reaches critical mass

HCI has also moved out of the shadows. We recently heard news thet HCI is outselling CI. Howard and I attribute this to the advances made in VMware’s vSAN 6.2 and the appliance-ification of HCI. That and we suppose NVMe SSDs (see above).

HCI makes an awful lot of sense for application clusters that VMware is touting these days. CI was easy but an HCI appliance cluster is much, simpler to deploy and manage

For VMware HCI, vSAN Ready Nodes are available from just about any server vendor in existence. With ready nodes, VARs and distributors can offer an HCI appliance in the channel, just like the majors. Yes, it’s not the same as a vendor supplied appliance, doesn’t have the same level of software or service integration, but it’s enough.

[If you want to learn more, Howard’s is doing a series of deep dive webinars/classes on HCI as part of his friend’s Ivan’s ipSpace.net. The 1st 2hr session was recorded 11 December, part 2 goes live 22 January, and the final installment on 5 February. The 1st session is available on demand to subscribers. Sign up here]

Computional storage finally makes sense

Howard and I 1st saw computational storage at FMS18 and we did a podcast with Scott Shadley of NGD systems. Computational storage is an SSD with spare ARM cores and DRAM that can be used to run any storage intensive, Linux application or Docker container.

Because it’s running in the SSD, it has (even faster than NVMe) lightening fast access to all the data on the SSD. Indeed, And the with 10s to 1000s of computational storage SSDs in a rack, each with multiple ARM cores, means you can have many 1000s of cores available to perform your data intensive processing. Almost like GPUs only for IO access to storage (SPUs?).

We tried this at one vendor in the 90s, executing some database and backup services outboard but it never took off. Then in the last couple of years (Dell) EMC had some VM services that you could run on their midrange systems. But that didn’t seem to take off either.

The computational storage we’ve seen all run Linux. And with todays data intensive applications coming from everywhere these days, and all the spare processing power in SSDs, it might finally make sense.

Futures

Finally, we turned to what we see coming in 2019. Howard was at an Intel Analyst event where they discussed Optane DIMMs. Our last podcast of 2018 was with Brian Bulkowski of Aerospike who discussed what Optane DIMMs will mean for high performance database systems and just about any memory intensive server application. For example, affordable, 6TB memory servers will be coming out shortly. What you can do with 6TB of memory is another question….

Howard Marks, Founder and Chief Scientist, DeepStorage

Howard Marks is the Founder and Chief Scientist of DeepStorage, a prominent blogger at Deep Storage Blog and can be found on twitter @DeepStorageNet.

Raymond Lucchesi, Founder and President, Silverton Consulting

Ray Lucchesi is the President and Founder of Silverton Consulting, a prominent blogger at RayOnStorage.com, and can be found on twitter @RayLucchesi. Signup for SCI’s free, monthly e-newsletter here.