Upload
nguyenthuan
View
212
Download
0
Embed Size (px)
Citation preview
Bringing Intelligence to StorageProducing Efficient Scalable Systems
Intelligent Storage Solutions
Dr. Jae Young Do Microsoft Research
Dr. Vladimir Alves CTO & Co-Founder
Scott ShadleyVP Marketing
BEST OF SHOW WINNER
A Real World Problem – Finding A Lost football fanNeedle in a Haystack
August 9th, 2018 NGD Systems, Inc - Keynote Presentation FMS 20182
Wasting Time Moving Data
Need a smarter way how we
use storage to search data
Facing the men is an immense display with 16 screens. It conveys live images from network of 180 CCTV cameras
Today’s way of finding that needle…
August 9th, 2018 NGD Systems, Inc - Keynote Presentation FMS 20183
Survey Question:
Where do you see the
most storage performance
bottlenecks?
Source: ATM Research
THE Storage!
August 9th, 2018NGD Systems, Inc - Keynote Presentation FMS 2018
4
3 Factors driving the Storage Problem… Tsunami
‘Tsunami of data’ could consume one fifth of global electricity by 2025
August 9th, 2018 NGD Systems, Inc - Keynote Presentation FMS 20185
3 Factors driving the Storage Problem… Power
A heat-exchange process commonly used for cooling
submarines to the underwater datacenter. The system
pipes seawater directly through the radiators on the back
of each of the 12 server racks and back out into the ocean
Free coolingThere are quite a few data centers that have embraced "free cooling" totally, i.e. using the cold air outside.All you need is ... a mild climate
Gartner Says Data Center Power, Cooling and Space
Issues Are Set to Increase Rapidly as a Result of New
High-Density Infrastructure DeploymentsSTAMFORD, Conn., May 13, 2010
August 9th, 2018 NGD Systems, Inc - Keynote Presentation FMS 20186
3 Factors driving the Storage Problem… Near-Data ComputePUSHED TO THE EDGEFebruary 19, 2018 Timothy Prickett Morgan
Three motivating factors for using Edge Computing1. Preserve privacy
2. Reduce latency3. Be robust to connectivity issues
Near-Data Computation: Looking Beyond BandwidthPublished in: IEEE Micro ( Volume: 34, Issue: 4, July-Aug. 2014 )
August 9th, 2018 NGD Systems, Inc - Keynote Presentation FMS 20187
Solving the Data Growth and Compute Problem
August 9th, 2018 NGD Systems, Inc - Keynote Presentation FMS 20188
Challenges with Moving (Big) Data
bandwidth mismatchby >60X
power density
volumetric density
data bottleneck
August 9th, 2018 NGD Systems, Inc - Keynote Presentation FMS 20189
Using to Tackle Data Bottlenecks
Brings harmony backto bandwidth needs
Seamless Programming Model
Manage Capacity Growth
Scalability
August 9th, 2018 NGD Systems, Inc - Keynote Presentation FMS 201810
Dimensions that Enable Computational Storage
32-bit real-time processors
AI acceleration
hardware
firmware
container virtualization
application software
RTOS
64-bit OS
bare metal
AI applications
operating system
user application
August 9th, 2018 NGD Systems, Inc - Keynote Presentation FMS 201811
In-Situ Processing Ecosystem – The Newport Platform
14nm
FinFet
August 9th, 2018 NGD Systems, Inc - Keynote Presentation FMS 201812
Delivering the Solution – NGD Systems NVMe SSD Family
8TB / 8W 32TB / 12W
64TB / 13WUp to 64TB
16 flash channels16TB / 12W
New Rack-Scale Form Factors Traditional Storage Form Factors
August 9th, 2018 NGD Systems, Inc - Keynote Presentation FMS 201813
Use Cases
Object Tracking – Azure IoT Hub
AI Image Classification
August 9th, 2018 NGD Systems, Inc - Keynote Presentation FMS 201814
Introducing Jae Young Do –Finding the Right Path Forward
August 9th, 2018 NGD Systems, Inc - Keynote Presentation FMS 201815
Case Study: Conventional SSD Storage Server (1/2)
Flash SSD
Flash SSD
Flash SSD
Flash SSD
Flash SSD
Flash SSD
Flash SSD
Flash SSD
Ro
ot
com
ple
xCPU
DR
AM
PCIeSwitch
PCIeSwitch
PCIeSwitch
Easy to scale up
High cost of moving data
August 9th, 2018 NGD Systems, Inc - Keynote Presentation FMS 201816
CPU
DR
AM
Ro
ot
com
ple
x Flash SSD
Flash SSD
Flash SSD
Flash SSD
Flash SSD
Flash SSD
Flash SSD
Flash SSD
PCIeSwitch
PCIeSwitch
PCIeSwitch
1
2
32
1
2
32
16
2
11
2
3
4
1
2
3
4
32 channels X ~500 MB/s = ~16 GB/s
64 Flash SSDs X ~16 GB/s/SSD = ~1TB/s
16 lanes of PCIe = ~16 GB/s
Throughput gap
of 66x!
Case Study: Conventional SSD Storage Server (2/2)
August 9th, 2018 NGD Systems, Inc - Keynote Presentation FMS 201817
Programming Attempts with standard SSDs
Not dev-friendlyprogramming environment
Not accessibleprototype devices
analytic query
Scan
Filter
Report
Not enough spareprocessing power
August 9th, 2018 NGD Systems, Inc - Keynote Presentation FMS 201818
Disruptive Trends that Enable Intelligent SSDs
Frugal resources inside SSD
Abundant resources inside SSD
Embedded CPU, proprietary firmware
General purpose CPU, server-like OS
(CP
U #
core
s/cl
ock
sp
eed
, h
ard
war
e o
fflo
ad,
DR
AM
)
(Ease of programmability inside SSD)
Today’s SSD
Intelligent SSD
August 9th, 2018 NGD Systems, Inc - Keynote Presentation FMS 201819
CPU
DR
AM
Ro
ot
com
ple
x Flash SSD
Flash SSD
Flash SSD
Flash SSD
Flash SSD
Flash SSD
Flash SSD
Flash SSD
PCIeSwitch
PCIeSwitch
PCIeSwitch
1
2
32
1
2
32
16
2
11
2
3
4
1
2
3
4
64 Flash SSDs X ~16 GB/s/SSD = ~1TB/s
16 lanes of PCIe = ~16 GB/s
Throughput gap of
66x!
~2GHz/core X 8 cores X 64 SSDs= ~1THz
~2GHz/core X 20 cores= ~40GHz
Compute capability
gap of 25x!
Revisit: Conventional SSD Storage Server
August 9th, 2018 NGD Systems, Inc - Keynote Presentation FMS 201820
The SoftFlash project proposes to create a software-defined storage substrate of flash SSDs in the data center that is as programmable, agile, and flexible as the applications and operating systems accessing it from servers.
– Embrace flash SSDs as a first-class programmable platform in the cloud data center
– Add custom capabilities to storage over time– Better bridge the gap between application needs and flash media
capabilities/limitations– Innovate in-house at cloud speed
MSR SoftFlash Project
August 9th, 2018 NGD Systems, Inc - Keynote Presentation FMS 201821
Memory
Related Images
0
2000
4000
6000
0.1B 0.5B 1B 5B 10B
GB
# Images
Required Memory Space~4TB DRAM!!
Example Scenario with Intelligent SSDs – Image Search
What if we had
Intelligent SSDs?
August 9th, 2018 NGD Systems, Inc - Keynote Presentation FMS 201822
0
500
1000
1500
2000
Host-only w/ 5 iSSDs w/ 10 iSSDs w/ 15 iSSDs
Qu
eri
es p
er
Seco
nd
Preliminary Results: Image Query Throughput
NOTEThe I/O cost of moving data is NOT considered! More interesting results are …
~4X Throughput Improvement
August 9th, 2018 NGD Systems, Inc - Keynote Presentation FMS 201823
• Finding the Needle FasterOn Drive Linux OS, Container Support
Dedicated Compute Cores
• Bigger Pipes Feed Smaller OnesMitigating Data Movement
Optimizing Application Execution
• Smarter Storage Does WorkPartnerships for Success
Real World Implementation
• Requires Intelligent ControllersFlash Agnostic – ONFI/Toggle, TLC/QLC16 Channels - Capacities to 256TB
• Power is Factor - Always.35 W/TB @ 16TB
Key Takeaways NGD Systems Newport Platform Provides
August 9th, 2018 NGD Systems, Inc - Keynote Presentation FMS 201824
Bringing Intelligence to Storage
Thank You
BEST OF SHOW WINNER