Monthly Archives: May 2019

EP24 – Pavilion Data: NVMe-oF for Modern Scale-Out Applications and Massively Parallel Compute – with Walt Hinton

In this 24th episode of the TECHunplugged Podcast we welcome Walter Hinton, Head of Corporate and Product Marketing at Pavilion Data. This episode was recorded live at Dell Technologies World 2019 in Las Vegas.

Podcast co-host Max Mortillaro (@darkkavenger) talks with Walt about the challenges with traditional all-flash infrastructures and the specific needs of modern scale-out applications. Walt covers extensively those topic, then goes on to explain why NVMe-based systems need purpose-built architectures, and how Pavilion Data is a great fit for scale-out / massively parallel applications or filesystems.

About Pavilion Data

Pavilion Data is the industry’s leading NVMe-oF Storage Array. It is a true end-to-end NVMe solution, from the host all the way down to the media. The Array is 100% standards compliant with zero host-side presence and was designed for the modern massively parallel clustered web and analytics applications.

Pavilion’s Storage Array delivers the next generation of composable disaggregated infrastructure (CDI) by separating storage from computing resources to allow them to scale and grow independently. In today’s large-scale environments it allows customers to become more agile by delivering the exact amount of composable resources at any given time.

Top view of a Pavilion Data system showing the NVMe drives, the PCIe Fabrics, and the high speed ethernet interconnects – the rack size of a system is 4U

About Walt

Walt is responsible for Corporate and Product Marketing.  He brings a deep technical background along with proven success in building marketing teams and programs.  Over a 25-year career in data storage, Walt has helped build successful startups like McDATA, ManagedStorage International and Virident Systems. He also served as Chief Strategist at StorageTek where he was instrumental in the creation of the Storage Network Industry Association (SNIA).  Most recently, Walt was Sr. Global Director of Product Marketing at Western Digital. He has a BA from William Jewell College and an MBA from the University of Denver.

Show schedule:

  • 00:00 Introduction, Walt’s Presentation
  • 02:00 Pavilion Data in numbers
  • 02:23 Traditional all-flash array architectures aren’t designed for NVMe because of bottleneck around controllers
  • 03:25 Rebuild times in case of a node loss (cca 25 minutes per TB) put a limit to the node capacity of direct attached storage in scale-out storage architectures
  • 05:25 Pavilion Data: A storage design inherently built for NVMe: scalable controllers, plenty of networking (40x 100 GbE), a switch-based PCIe backplane, and the ability for customers to source the NVMe drives of their choice
  • 06:40 Walt explains that Pavilion Data’s architecture allows for a data rebuild at a rate of 5 minutes per TB
  • 07:32 Use cases, industries and verticals for Pavilion Data
  • 08:22 The perfect fit for Pavilion Data: scale-out applications leveraging Cassandra, MongoDB, MariaDB etc.
  • 08:38 The Pavilion Data – Pivotal partnership – supporting Greenplum (an open-source massively parallel data platform for analytics, machine learning and AI)
  • 09:01 A take on financial services, massively distributed databases, and backup challenges with multi-petabyte data lakes
  • 10:20 Talking about protocols (Pavilion Data is block-based) and clustered filesystems (Spectrum Scale, etc.)
  • 11:41 Continuing the discussion on supercomputing and massively parallel compute, media & entertainment, as well as government
  • 13:05 Describing the physical aspects of a Pavilion Data system
  • 13:50 A 4U, fully fault-tolerant system achieving 120 Gb/s reads or 90 Gb/s writes – what is the equivalent with a traditional AFA?
  • 15:05 The metaphor of the nail and the hammer
  • 15:31 Partnerships & Sales – how to engage Pavilion Data
  • 16:25 The partnership with Dell
  • 17:30 The synergy between Pivotal and Pavilion Data – embracing customer needs
  • 19:15 Talking about worldwide availability
  • 20:28 Closing remarks

EP23 – Lightbits Labs: NVMe Flash Performance Commoditization via NVMe over TCP – with Kam Eshghi

In this 23rd episode of the TECHunplugged Podcast we welcome Kam Eshghi, VP of Strategy & Business Development at Lightbits Labs. This episode was recorded live at Dell Technologies World 2019 (early May 2019) in Las Vegas.

Lighbits Labs have developed a software-defined storage solution leveraging NVMe over TCP. Their solution allows the disaggregation of storage from compute by offering DAS performance with enterprise-class data services, combined with massive scalability.

A sizeable part of the founding team was behind DSSD, a storage solution which was later purchased by Dell EMC. DSSD was one of the very first storage architectures leveraging NVMe drives, which gives some confidence about Lightbits Labs NVMe over TCP concept. Lightbits Labs was founded 3 years ago and went out of stealth mode in April 2019, so we’re really thrilled to get prime time with them!

During the course of the discussions, podcast co-host Max Mortillaro (@darkkavenger) talks with Kam about the product architecture, its concepts and the use cases for NVMe over TCP.

Show schedule:

  • 00:00 Presentations, introduction to Lightbits Labs and NVMe over TCP
  • 02:00 Addressing the challenge of scalability of DAS and performance of traditional architectures
  • 03:30 The genesis of NVMe over TCP, and why NVMe over TCP is relevant today
  • 05:20 Kam states that no specific drivers are needed, and that Lightbits Labs made its source code available
  • 05:45 All of the intellectual property of Lightbits Labs resides on the target side; Kam also mentions that Lightbits Labs sell a hardware solution called SuperSSD, and that they also have an optional accelerator card
  • 06:22 Max & Kam discuss about partnerships & go to market strategy: software-only, via Dell OEM, or via the Lightbits SuperSSD appliance
  • 08:00 Kam: « All you need to get started is a server with NVMe drives and a standard Ethernet NIC »
  • 08:30 Let’s talk architecture and data services
  • 09:45 Kam mentions a « Global FTL » – that has Max interested in understanding how the internal logic of NVMe SSD drives is managed
  • 12:30 More insights into data services
  • 13:45 Understanding the customer base and use cases for Lightbits Labs: SaaS companies, Service Providers, etc.
  • 15:40 Talking about data protection, replication, and availability
  • 16:30 Application use case: distributed databases (Cassandra, MongoDB, etc.) and High-Performance Analytics workloads, and over all anything that requires high performance and operates at scale
  • 18:00 Max’s usual « WOW » / speechless moment; Kam shares his excitement about adopting customers, not only Hyperscalers but also Private Cloud initiatives in with Enterprise IT organisations
  • 19:00 Since Lightbits Labs is block-based, Max raises the question as to whether there are any plans to offer managed services in the plan, if it makes sense at all
  • 20:15 Covering the topic of licensing models
  • 21:09 What about the optional acceleration card? Kam explains that the card sits on storage nodes, and the decision points about why it may make sense to use the acceleration card. This offers flexibility to customers who may want to select entry level CPUs to keep costs in control.
  • 23:30 Thanks and Conclusion

EP22 – Excelero – NVMe Software-Defined Storage on Steroids – with Josh Goldenhar

In this 22nd episode of the TECHunplugged Podcast we welcome Josh Goldenhar, VP of Products at Excelero. Prior to joining Excelero, Josh was responsible for product strategy and management at EMC (XtremIO) and DataDirect Networks.

Podcast co-hosts Max Mortillaro (@darkkavenger) and Arjan Timmerman (@Arjantim) talk with Josh about Excelero, the solution’s architecture, its use cases & differentiators.

Show schedule:

  • 00:00 Presentations & introduction to Excelero
  • 01:50 We learn that Excelero is software-based and uses NVMe drives. We ask Josh about whether there is a Hardware Compatibility List. Josh goes on to talk about custom built vs. co-engineered solutions (Dell, SuperMicro) and mentions a recently announced partnership with Lenovo.
  • 03:25 Josh explains how Excelero was built from the start to be hardware agnostic and provides a perspective about how each hardware vendor is looked at based on their own hardware specificities.
  • 04:46 Consuming Excelero: what customers need to do to get their Excelero storage up and running: either via pre-built appliances or via custom built hardware with installation of a couple RPM packages.
  • 06:40 Is Excelero block-based or file-based? Josh explains that Excelero is a block-based distributed storage and provides background about the rationale to go block-based only.
  • 08:40 We ask Josh about Excelero’s customers and their use cases. Low latency, consistency in response times, and ability to scale are key to those customers. Josh then goes on to explain some of the common challenges faced by web-scalers and how Excelero fits in the picture.
  • 11:25 The case of Technicolor, an Excelero customer – how the motion picture industry requires bandwidths akin to those used in HPC clusters
  • 12:35 Excelero storage deployment modes (disaggregated vs. converged) and their technical implications
  • 15:35 A look into network interconnects that are supported by Excelero and throughput capabilities
  • 19:20 Talking about HPC and Local Burst Buffer / Local Scratch – integration with SLURM job scheduler and local nodes
  • 23:50 Being mind blown and forgetting a question – the Venn Diagram of Happy and Sad
  • 24:20 Remembering a question – does Excelero supports different media / performance tiers (such as SATA SSD or HDD), or 3D XPoint. Josh provides a comprehensive view of what is supported backed by Excelero’s rationale about why things are being done or implemented in a specific way.
  • 27:30 How / where to purchase Excelero, and how is it licensed?
  • 30:15 Conclusion