Sponsored By - G2M Incorporated...

RESEARCH

G2M Research Webinar onComputational Storage andEDSFF for Flash DevicesNovember 13, 2018

Sponsored By:

2 © 2018 G2M Communications, Inc.. All rights reserved.RESEARCH

Scott ShadleyVice President,Marketingwww.ngdsystems.com

Panelists

Mike HeumannManaging Director

www.g2minc.com/research

Jonathan HinklePrincipal ResearcherSystems Architecturewww.lenovo.com

Mike ScriberProduct Director

www.supermicro.com

http://www.solarflare.com/

http://www.mellanox.com/

http://www.broadcom.com/

http://www.broadcom.com/


Webinar Agenda

9:00-9:09 Introduction and NVMe SSD Market Overview(Mike Heumann, G2M Research)

9:10-9:23 Computational Storage Solutions (Scott Shadley, NGD)

9:24-9:24 Audience Survey 1 (1 minute)

9:25-9:38 Petabyte-Scale Platform Solutions (Mike Scriber, Supermicro)


9:40-9:53 New SSD Form Factors (Jonathan Hinkle, Lenovo)


9:54-9:58 Audience Q&A (8 minutes)

9:59-10:00 Wrap-Up

RESEARCH

Introduction and NVMe SSDMarket OverviewMike HeumannManaging PartnerG2M Researchwww.g2minc.com/research


What is a Flash Solid State Drive (SSD)?

A simple view of SSDs is that they are a bunch of NAND chips on a board, mounted in a protective carrier

They can have more or less buffering RAM

Some “driver” or management capabilities can also be embedded in the controller

AND they have “standard” PC formats (2.5” drive or PCIe add-in card formats)


How Have Flash SSDs Changed The Enterprise Data Center?

Prior to flash, storage was the “long pole” in the IT performance tent– That is why enterprise storage arrays were so expensive– The arrays were chocked full of DRAM– And they contained the fastest HDDs that money could buy

Flash storage has changed the performance narrative– Storage latencies went from milliseconds to microseconds– Non-Volatile Memory Express (NVMe) reduced it even further

Now the network and the PCIe bus are the long pole– In many data centers, moving data takes more time and

as much power as processors do


The Challenges With Enterprise SSDs Today

SSD capacity is increasing, but it is still form-factor bound– PCIe cards have capacity, but have lots of other issues– M.2 are low-cost, but also low-capacity and not hot-swappable– U.2 is not optimized for NAND flash chip layout, airflow

So far, U.2 is the clear winner in the NVMe flash race– 3.7M units shipped in 2017 (over 4X what shipped in 2016)– 29 models, covering nearly every major storage vendor

M.2 also showing good growth (but not as strong as U.2)– Shipped 1.95M units in 2017(roughly 50% growth from 2016)– 25 models, mostly from leading storage vendors

PCIe add-in Cards are roughly flat (224K units in 2017)– Less models than in 2016


What Does the Future Hold for Enterprise SSDs?

The new Enterprise and Datacenter SSD Form Factor (EDSFF) is one new option– EDSFF is tailored for flash media (better airflow,

better layout for higher density)– EDSFF is (or can be) hot-swappable– EDSFF comes in multiple form factors

Companies are also adding new capabilities to SSDs – Computational storage– Network/fabric interfaces– Increased redundancy/error correction– More than just adding additional capacity or

performance

Bringing Intelligence to Storage


The Future of Storge for 2019+

October 2018


Computational Storage Solutions

Nov 13th, 2018 NGD Systems, Inc. - Presented Nov 2018


Avoiding Clogged Pipes!! Moving Compute to Data is Cheaper!!• A compute request is more efficient if executed @ the data

– Minimizes network traffic from Edge to Fog to Cloud– Increases throughput & performance of the system – Enables distributed processing– Especially true for Big Data (analytics):

• Traditional approach: move ALL the data for search– Eventually limited by networking bottlenecks

• New approach: Search the data in place, deliver the RESULTS– Give back CPU resources and Bandwidth


Presenter

Presentation Notes

Example: HDFS provides interfaces for applications to move computation closer to where data is located Moving computation to data is better than the other way around


Challenges with Moving (Big) Data


Host Platform

Standard NVMe Protocol


Standard SSDIntelligent

STD Controller

Key Attributes: Maintain familiar methodology

Linux, APIs, SDK, DockerUse standard protocols (NVMe) and processes

No new commandsMinimize interface traffic

Power and time savingsEnhancing limited footprint with maximum benefit

Customer TCO

Presenter

Presentation Notes

Thanks for the introduction, Scott. We all know that data is growing at an enormously fast pace and it is also becoming invaluable for businesses and society in general. Data needs to be analyzed, parsed and mined. The increasing overheads of computing-centric architectures <CLICK> exemplified by the storage to computing bandwidth mismatch make near data processing an attractive alternative for at least two reasons: - moving computation close to data reduces data movement; - and providing a better matching of computation to data capacity, bandwidth, and locality makes for more balanced systems. On the software side, distributed software frameworks such as MapReduce and Hadoop have popularized both the concept of moving computation to data as well as programming near data processing hardware. To make it a reality we need to tackle at least 3 important factors related to storage and computation: <CLICK> Power density – measured in watts per terabyte <CLICK> Volumetric density – measured in terabytes per cubic centimeter And, one of the biggest headaches for computing systems: <CLICK> data bottlenecks


Computational Storage in the Real World – Little Fish, Big Pond

TODAY – Catch and Release, Over and Over… FUTURE - NGD – Fish Jumps Out of the Lake!

Today’s Datacenters have massive sets of data, much like schools of fish.

How do you find the one Orange fish?

You catch them all one at a time and release, then repeat after seeing each one … Not Practical Right?

What if that fish, shown here, simply JUMPS out of the water…

This method of pre-screening and locating only the fish, or data, needed is the premise of Computational Storage…

Finding NEMO has a whole new meaning… Especially when today’s datacenter Servers have a lot of DORYs!


Host Managed FTL Adds power, DRAM, & Data traffic

“Computational Storage” Paths – Shorter is Always BetterFPGA Only for

Compute and NAND –Cost, Scale Issues

FPGA/SSD Controller adds Data Distance,

Memory Cost

FPGA to Multi-SSD Adds Size, More Data Distance, Power, Cost

SINGLE ASIC Solution Lowest Power, Cost

Shortest Data Distance



In-SituSoftware

Optimized Hardware

SiliconArchitecture

Patented Firmware

FlexibleManagement

Algorithms

Computational Storage PlatformPA

TEN

TED

NGD

Sys

tem

s Sol

utio

ns

EDSFF/M. 2Up to 16TB

Quad-core 64-bit application processor

Light Virtualization

In-Situ Software Defined Hardware acceleration

Full Fledgedon drive OS

U.2 15mm Gen3 (x4)Up to 32TB

AIC Gen3 (x4)Up to 64TB

ProprietaryNVMe Controller



SNIA to Develop “Standards” for Computational Storage




The Future of Storage is Now!Thank You

You can find us at:

www.NGDSystems.com Twitter: @NGDSystems Facebook: @NGDSystems

Scott Shadley – [email protected] - @SMShadley


http://www.ngdsystems.com/

https://twitter.com/NGDSystems

https://www.facebook.com/NGDSystems

mailto:[email protected]


Audience Survey Question #1

Does standardizing computational storage via SNIA make you more likely to adopt it within the next 12-24 months? (choose one):

• We would adopt computational storage in the next 12-24 monthseven if SNIA didn’t provide standardization for the industry 0%

• Definitely – very important to my organization’s decision-making process on computational storage 32%

• It helps, but it is only one of the factors that we look at 44%

• No impact – we would make our decisions regardless ofSNIA standardization efforts 4%

• Not plans to utilize computational storage in the next 12-24 months 4%

• Not familiar with SNIA’s computational storage standardization efforts 4%

• Who is SNIA? 12%

© Supermicro 2018

Confidential

We Keep IT Green™

New Directions in PB-Scale Hardware Solutions

Product Director / Mike Scriber

11/13/2018

A PetaByte in a 1U

© Supermicro 2018

Confidential

PCI Express* SSD Intel Form Factor Evolution

M.2

U.2

AIC

Ruler

M.2 initially designed for client and mobile use. Used in data center for boot or compute nodes, but lacks hot-plug support and requires carrier

cards / heatsinks to manage thermals

U.2 2.5in x 15mm and 7mm supports hot-plug and serviceability, designed to share physical

dimensions with HDDs for hybrid HDD/SSD server designs. Mainstream PCIe* SSD form factor

PCIe* low profile add-in-cards have broadest compatibility with the most mature ecosystem and

compliance. Shares same form factor with network cards, graphic cards, etc.

Built for data center racksHigh per drive, per server and per rack capacity

Improved manageability and serviceabilityEfficient thermal design

Integrated enclosure, latch, LEDs

*Other names and brands may be claimed as the property of others.

© Supermicro 2018

Confidential

High Density & High Capacity: 1U 32 hot-swap NVMe drives x 32TB = 1 PBServiceability: Hot-swap fan modules and power supplies Hot-swap SSDs & Tool-Less Drive TraysAccessibility: Remote system on/off and system management Remote power cycling for each Individual SSD Redundancy: Redundant Power Supplies and FansPerformance: Flexible to configure up to 8 Hosts High throughput up to 64 TB/s data transfer rate Balanced design with scalability at core

Product Requirements

© Supermicro 2018

Confidential

Supermicro Confidential

Hot-Swap Modular Design

32 Hot-Swap U.2 Drives

2 Hot-Swap Redundant Power Supplies8 Hot-Swap Fan modules

© Supermicro 2018

Confidential

Tool-less Drive Tray Design

Umbilical Cable

Tool-less Drive Carrier

© Supermicro 2018

Confidential

Application Scenarios

High capacity storage requirements

High Throughput Ingest High Density Hot Storage HPC / Data Analytics Media/Video Streaming Content Delivery Network (CDN) Big Data Top of Rack Storage

© Supermicro 2018

Confidential

1U 32 U.2 NVMe Server

Supermicro Confidential

2 Skylake Processors24 DIMMs

2 X16 PCIe Slots

32 Hot-Swap U.2 Drives

© Supermicro 2018

Confidential

Intel Ruler Advantages

25

• A x4 Orthogonal Connector Configuration• Tips the card edge up to accept a front

mounted PCB at 90° to the system PCB• Air naturally flows through the connector for excellent

thermals• Connectors “Gang” together in a frame that screw-

attaches to the PCB for excellent mechanical integrity• Presents the X4 PCIe* lanes in a very routing friendly

manner• Through-hole and supports PCIe* gen 4 speeds

today, surface mount to support gen 5• Reduces cost, improves reliability, and allows for

unimpeded airflow

© Supermicro 2018

Confidential

Ruler JBOF and Server

• Fully front Serviceable -Fits into 1U• Maximize drive capacity• Maximize capacity per 1U• Optimize thermal efficiency• Drive down TCO

Why Ruler SSD?

1U 32 Ruler JBOFSSG-136R-NR32R

• Up to 8 host connections• Optional NVMe over

Fabric using Attala AOC• HPC / Data Analytics • RSD/Top of Rack

Storage

1U 32 Ruler ServerSYS-1029P-NR32R

• DP w/ 24DIMM• High Throughput Ingest• HPC / Data Analytics • Content Delivery Network • Top of Rack Storage

© Supermicro 2018

Confidential

Samsung NF1 Server

© Supermicro 2018

Confidential

SSG-1029P-NMR36L

© Supermicro 2018

Confidential

Thank YouFor more information, visit www.supermicro.com or contact [email protected]

We Keep IT Green™

© Supermicro 2018

Confidential

Disclaimer Super Micro Computer, Inc. may make changes to specifications and product descriptions at any time, without notice. The information presented in this document is for informational purposes only and may contain technical inaccuracies, omissions and typographical errors. Any performance tests and ratings are measured using systems that reflect the approximate performance of Super Micro Computer, Inc. products as measured by those tests. Any differences in software or hardware configuration may affect actual performance, and Super Micro Computer, Inc. does not control the design or implementation of third party benchmarks or websites referenced in this document. The information contained herein is subject to change and may be rendered inaccurate for many reasons, including but not limited to any changes in product and/or roadmap, component and hardware revision changes, new model and/or product releases, software changes, firmware changes, or the like. Super Micro Computer, Inc. assumes no obligation to update or otherwise correct or revise this information.

SUPER MICRO COMPUTER, INC. MAKES NO REPRESENTATIONS OR WARRANTIES WITH RESPECT TO THE CONTENTS HEREOF AND ASSUMES NO RESPONSIBILITY FOR ANY INACCURACIES, ERRORS OR OMISSIONS THAT MAY APPEAR IN THIS INFORMATION.

SUPER MICRO COMPUTER, INC. SPECIFICALLY DISCLAIMS ANY IMPLIED WARRANTIES OF MERCHANTABILITY OR FITNESS FOR ANY PARTICULAR PURPOSE. IN NO EVENT WILL SUPER MICRO COMPUTER, INC. BE LIABLE TO ANY PERSON FOR ANY DIRECT, INDIRECT, SPECIAL OR OTHER CONSEQUENTIAL DAMAGES ARISING FROM THE USE OF ANY INFORMATION CONTAINED HEREIN, EVEN IF SUPER MICRO COMPUTER, Inc. IS EXPRESSLY ADVISED OF THE POSSIBILITY OF SUCH DAMAGES.

ATTRIBUTION© 2018 Super Micro Computer, Inc. All rights reserved.



How much total capacity do you require for each all-flash system(choose one)?

• 64TB-128TB: 8%

• 128TB-256TB: 0%

• 256TB-512TB: 8%

• 512TB-1PB: 42%

• 1PB-8PB: 23%

• 8PB-16PB: 4%

• More than 16PB 15%


Datacenter server systems

Many different types of systems, but most servers today: rack-mount and fit in multiple of the “1U” rack space.

Mission Critical

Blade and ½ wide Compute Nodes

1U rack server


Datacenter SSD – Our Previous OptionsAdapter Cards HDD form factorsConsumer SSD

Good: High-perf CacheBad: Steal PCIe slots from IOUgly: $$$, You might fit two?

Good: Small and ModularBad: Low capacity, no hot-plug, connector less reliableUgly: More expensive once adapted to enterprise needs

Good: Hot-plug, Storage featuresBad: Designed for HDD, not SSDUgly: Blocks airflow to the hottest components in server

Need an optimized common building block across server designs.

35

Enter EDSFF

General purpose scalable connector Flexible: Multiple orientations, widths, PCIe* Gen 5+ support Supports interoperable specs (EDSFF NVMe, CCIX, OCP Mezz,

Gen Z, OpenCAPI, Ethernet, etc.)

Break free of legacy to optimize for NVM 50-100% increase in media package sites

Improved thermal efficiency 2-3x less airflow needed Support higher power devices

Industry standardConnector: SFF-TA-1002

https://www.amphenol-icc.com/product-series/mini-cool-edge-0-60mm.htmlSource – Intel. Comparing airflow required to maintain equivalent temperature of a 4TB U.2 15mm Intel® SSD DC P4500 to a 4TB “1U.L” form factor for Intel® SSD DC P4500. Source – Intel. Comparing airflow required to maintain equivalent temperature of an 8TB U.2 7mm Intel® SSD DC P4500 to a 8TB EDSFF 1U-Short form factor for Intel® SSD DC P4510.Results have been estimated or simulated using internal analysis or architecture simulation or modeling, and provided for informational purposes. Simulation involves three drives for each form factor in a sheet metal representation of a server, 12.5mm pitch for “Ruler” form factor, 1000m elevation, limiting SSD on case temp of 70C or thermal throttling performance, whichever comes first. 5C guard band. Results used as a proxy for airflow anticipated on EDSFF spec compliant “Ruler” form factor Intel® SSD P4510.

E1.S

E1.L

U.2 7mm

U.2 15mm

(shown at FMS2018)Santa Clara, CA - August 2018

https://www.amphenol-icc.com/product-series/mini-cool-edge-0-60mm.html

36

EDSFF Scalable Family for Different Usages

E1.S (SFF-TA-1006)• 111.5 x 31.5 mm• Supports >12W• Up to 12 Standard NAND sites

E3 (SFF-TA-1008)• (104.9/142.2) x 76mm• Supports up to 70W• Up to 48 Standard NAND sites

E1.L (SFF-TA-1007)• 318.75 x 38.4 mm• Supports > 40W• Up to 48 Standard NAND sites

Same Protocol: NVMe Same Interface: PCIe Same Connector: SFF-TA-1002 Same Pinout and Functions (hot plug,

serviceable) Different Usages, Same Expectations!

Industry standards:


EDSFF 1U Short: the SFF-TA-1006 industry standardVision:Create a smaller, high density solid state drive standard that is optimized for the datacenter.

EDSFF 1U Short (E1.S):• Mainstream NVMe drive• Compact, modular form factor

• Vertical fit in 1U height• Fits in depth of 2.5” drive

• High Performance and Capacity• Supports hot-plug and enterprise feature set• +12V main power for reduced system cost• LEDs on-drive for lower cost and easier integration

Optimized for the NVMe drive design and use across all datacenter and edge systems to scale as mainstream storage.

Presenter

Presentation Notes

Mainstream NVMe drive – targeted to best fulfill the 80% sweet spot, but can extend to highest performance and similar capacity as 7mm u.2

38

Industry Standard datacenter-optimized NVMe drive that provides significant new system benefits

• Key benefits:‒ Much smaller enabling high density storage‒ Significantly improved system airflow and thermal solution‒ Most efficient modular scaling of NVMe capacity and performance Enhanced feature set in

space-constrained servers‒ Lower base system infrastructure and drive costs (high volume, common building block)

EDSFF 1U Short (E1.S) - Key System Benefits


Presenter

Presentation Notes

The E1.S provides significant new system benefits Key system benefits: Much smaller enabling high density (4X EDSFF 1U Short vs. 2.5” drives in 1U) Better modular scaling (finer granularity… customers can scale storage in smaller increments) Enables other system benefits: More complete and competitive feature set in space-constrained servers Opportunity to improve cost, power, and thermals by eliminating backplane Smooth SSD / HDD transition in servers as NVM prices continue to drop System build configuration option for adding support as desired 6-8 E1.S in space of 2 x 2.5” drives Much better airflow: 70% lower pressure drop across drives Flexibly scales performance and capacity in system with same high-volume drive Unlocks inherent performance of flash media on drive without expensive, high-end controllers

39

Build Your Future with EDSFF

EDSFF is THE form factor for the next decade for SSDs Higher power, higher bandwidth, better thermals and better density

Future proof you next design with the common connector SSDs, NICs, FPGAs, Persistent Memory, Accelerators, …

For more information: http://www.snia.org/sff/specifications Connector: SFF-TA-1002: Card Edge multilane protocol agnostic connector E1.S formfactor: SFF-TA-1006: Enterprise and Datacenter 1U Short SSD Form Factor E1.L formfactor: SFF-TA-1007: Enterprise and Datacenter 1U Long SSD Form Factor E3 formfactor: SFF-TA-1008: Enterprise and Datacenter 3” Media Device Form Factor Pin list/behavior: SFF-TA-1009: Enterprise and Datacenter SSD Pin and Signal Specification

http://www.snia.org/sff/specifications


Thank you!



What are the biggest challenges today in designing data center systems for NVMe flash storage? (check all that apply):

• Adding enough capacity: 27%

• Cost-effective scaling of performance: 68%

• Thermal challenges for drives and systems: 45%

• Adapting SSDs that were not meant for data center use: 23%

• Product issues relating to NVMe maturity: 27%

• Other 9%

• None 5%


Audience Q&A

RESEARCH

Thank You For Attending

Sponsored By - G2M Incorporated...

Documents

Transcript of Sponsored By - G2M Incorporated...