Upload
others
View
0
Download
0
Embed Size (px)
Citation preview
Moore’s Law and Big Data
Andreas Bechtolsheim
Chairman Arista NetworksCo-Founder DSSD
Agenda
• Moore’s Law
• Server Technology
• Storage Technology
• Big Data Architecture
• Use Cases
Moore’s Law is Alive and Well
Dr. Moore’s Prediction 1965
1965 Prediction: Density would double every year
10/29/09 12:49 PMhttp://upload.wikimedia.org/wikipedia/commons/0/00/Transistor_Count_and_Moore%27s_Law_-_2008.svg
Page 1 of 1
40 Y
1,000,000X
2X/2Y
Moore’s Law 1971-2011: 2X/2Y
Overall Roadmap Technology Characteristics —2008 Update 7
Average Industry
"Moores Law“ :
2x Functions/chip Per 2 Years
2007 ITRS Product Technology Trends -
Functions per Chip
1.E-02
1.E-01
1.E+00
1.E+01
1.E+02
1.E+03
1995 2000 2005 2010 2015 2020 2025
Year of Production
Pro
du
ct
Fu
nc
tio
ns
/Ch
ip
[ G
iga (
10^
9)
- b
its, tr
an
sis
tors
]
Flash Bits/Chip (Gbits)Multi-Level-Cell (4bitMLC)
Flash Bits/Chip (Gbits)Multi-Level-Cell (2bitMLC)
Flash Bits/Chip (Gbits)Single-Level-Cell (SLC )
DRAM Bits/Chip (Gbits)
MPU GTransistors/Chip- high-performance (hp)
MPU GTransistors/Chip- cost-performanc (cp)
2007 - 2022 ITRS Range
Average Industry
"Moores Law“ :
2x Functions/chip Per 2 Years
Average Industry
"Moores Law“ :
2x Functions/chip Per 2 Years
2007 ITRS Product Technology Trends -
Functions per Chip
1.E-02
1.E-01
1.E+00
1.E+01
1.E+02
1.E+03
1995 2000 2005 2010 2015 2020 2025
Year of Production
Pro
du
ct
Fu
nc
tio
ns
/Ch
ip
[ G
iga (
10^
9)
- b
its, tr
an
sis
tors
]
Flash Bits/Chip (Gbits)Multi-Level-Cell (4bitMLC)
Flash Bits/Chip (Gbits)Multi-Level-Cell (2bitMLC)
Flash Bits/Chip (Gbits)Single-Level-Cell (SLC )
DRAM Bits/Chip (Gbits)
MPU GTransistors/Chip- high-performance (hp)
MPU GTransistors/Chip- cost-performanc (cp)
2007 - 2022 ITRS Range
Figure ORTC2 ITRS Product Function Size Trends:
MPU Logic Gate Size (4-transistor); Memory Cell Size [SRAM (6-transistor); Flash (SLC and MLC), and
DRAM (transistor + capacitor)]--Updated
THE INTERNATIONAL TECHNOLOGY ROADMAP FOR SEMICONDUCTORS: 2008 UPDATE
100X
12Y
Silicon Roadmap 2012 - 2024
What can one do with 100 Billion Transistors per chip?
• Lot’s of CPU Cores • 100s of high-speed cores per CPU chip
• Lots of Memory Bits • PB Flash, Multi-TB DRAM Systems
• Lots of I/O and Network Bandwidth • 10s of Gigabytes/sec per CPU
Intel 4S Server Board (2015)
• 4 CPU Sockets
• 72 Cores
• 3+ GHz
• 48 DIMMs
• 3 TByte DRAM
• 1400W Power
SunFire E25K (2005)• 72 SPARC Cores
• 1.5 GHz
• 1.15 TByte DRAM
• 76”H, 33”W, 64”D
• 2900 lbs
• 28.7 KWatt
Less throughput than today’s single board 4S X86 Server
0
250
500
750
1000
2010 2012 2014 2016 2018 2020 2022
100X Performance Gain from 2012 to 2024
64-bit CPU Cores over Time
Average increase 30%/year
0
0.25
0.5
0.75
1
2010 2012 2014 2016 2018 2020 2022
Declining Cost of Computing
Average decline ~ 30%/year
• Moore’s Law 1971-2024 • Density has increased 2X every 2 Years • Million-fold improvement 1971-2011 • Another 100X expected 2012-2024 • 100 Million Fold Increase 1971-2024
• Moore’s Law is not ending, but slowing down • Economic and Physics limits on the horizon
Moore’s Law Summary
• Economics are getting more Challenging • Scaling is not “Free”
• Geometric Scaling is Hitting Physics Limits • 10nm, 7nm, 5nm, ???
• Certain Things Don’t Scale Well • For example Flash Bits
Moore’s Law Challenges
However there are always creative solutions to problems
Flash Scaling Issue• Flash does not scale well below 15 nm
• Not enough electrons left to store bits • Charge loss due to shallow trapped electrons • Reduced write and read endurance
• Solution: 3D Stacking • Older process technology with many layers • Enables 256 Gbit (MLC), 512 Gbit (TLC) • More electrons per bit improves endurance
Toshiba BiCS 3D-NAND Stack
DRAM Memory Scaling
© 2014 IBM Corporation4Flash Memory Summit 2015
DRAM & Flash Scaling – Density Trends
DRAM scaling at sub 20nm node for 12/16Gbit – bit cost reduction vs increased technology complexity & fab investment requirements
Flash scaling continues via 2D > 3D NAND transition – enabling 256Gb & higher density MLC/TLC Flash
DRAM Memory Scaling
© 2014 IBM Corporation4Flash Memory Summit 2015
DRAM & Flash Scaling – Density Trends
DRAM scaling at sub 20nm node for 12/16Gbit – bit cost reduction vs increased technology complexity & fab investment requirements
Flash scaling continues via 2D > 3D NAND transition – enabling 256Gb & higher density MLC/TLC Flash
Up to 16 3D NAND Die/Package
19
0
256
512
768
1024
2013 2014 2015 2016 2017
NAND Density/Package (GB)
2X
2X
2X
2X
Combination of TLC, 3D NAND, 16 Stack Die/Package
20
DRAM vs FLASH Pricing
Intel 3D X-Point NV Memory
• Read Performance
• Comparable to DRAM
• Write endurance • Much better than Flashbut not unlimited (10^7)
• Density
• Higher than DRAMbut lower than Flash
• Cost • Lower than DRAMbut higher than Flash
Intel 3DXP The Real Thing
Intel has researched PCM for 45 Years!
Intel 3X Xpoint Form Factors
DDR4 DIMM Formfactor Supports mixed DRAM/3DXP
Memory ConfigurationUp to 512 MB/DIMM
NVMe PCIe FormFactor Supports High-SpeedStorage Applications"Up to 4 TB/Container
How to Build Very High-performanceScalable Storage Compute Systems
The Memory Storage Hierarchy
“ Ideally one would desire an indefinitely large memory capacity such that any particular … word would be immediately available. … It does not seem possible physically to achieve such a capacity. We are therefore forcedto recognize the possibility of constructing a hierarchy of memories, each of which has greater capacity than the preceding but which is less quickly accessible.”
Preliminary Discussion of the Logical Design of an Electronic Computing Instrument Arthur Burks, Herman Goldstine and John von Neumann, 1946
Today’s Memory Storage Hierarchy
CPU Caches
DRAM / 3DXP
Flash Storage
Hard Disk
Optical / Tape
nanosecond Latency
10s of nanosecond Latency
Flash Storage
Disk
10 second Latency
100 microsecond Latency
10 millisecond Latency
• Data lives on Disk or Flash!• Deep Storage Hierarchy!• Apps use small % of data!• Legacy I/O Stacks (FC/SCSI)
• Data lives in persistent memory!• Many CPUs surround and use it!• Big Data Apps use entire datasets!• Need to rethink I/O Stack
Big DataCPU
PBs
TBsGBs
Memory
DiskTape
Old Compute-Centric Model New Data-centric Model
PetaByte High-Perf. Storage
CPU
CPUCPU
CPU CPU
CPU
CPU
CPU
Big Data Apps require rethinking of storage architecture
A New Data-centric Storage Model
Rack-scale DSSD Flash Storage
Copyright 2016 DSSD+EMC, Inc. Proprietary/Confidential: NDA Use Only
D5 Summary
22
36X 4T FMs (100TB usable)
96X Gen3 x4 Client Ports
Dual hot-swap controllers
Dual hot-swap I/O modules
Dual service modules
2+2 hot-swap power and fan
Flood Datapath
10 Million IOPS, 100 GByte/sec IO100 usec Latency, 100 TByte Storage
Rack-Scale Shared NVMe Storage
Copyright 2016 DSSD+EMC, Inc. Proprietary/Confidential: NDA Use Only
Rack with SSDs
3
Compute Node
SSDsControl+Data Traffic
Compute Node
SSDsControl+Data Traffic
Compute Node
SSDsControl+Data Traffic
Compute Node
SSDsControl+Data Traffic
Copyright 2016 DSSD+EMC, Inc. Proprietary/Confidential: NDA Use Only
Rack Scale Flash with DSSD
4
Compute
NodeControl Traffic
Compute
NodeControl Traffic
Compute
NodeControl Traffic
Compute
NodeControl Traffic
PCIe3 Data
PCIe3 Data
PCIe3 Data
PCIe3 Data
10/40Gb Uplink
10/40Gb Uplink
10M IOPS
100 GB/s
100 TB
100μs
Rack-scale NVMe Storage is bigger,faster, and more reliable than
local Flash Storage
NAND
System Call
Application
Libraries
Filesystem
Device Driver
( SAS/SATA )
PCIe HBA
Volume Manager
Flash Controller
NAND Flash
Kernel
Hardware
SoftwareApplication
Libraries
PCIe Link Card
DSSD Controller
DSSD Flash Module
5 ms H
DD
200µs SSD
20-100µs
25 µs OS
25 µs OS
User DMA Port
Rack-scale Direct Access Storage
100 usec Latency, All the Time
Copyright 2016 DSSD+EMC, Inc. Proprietary/Confidential: NDA Use Only
Hadoop I/O on HDD, SSD, and DSSD
11
DSSD Confidential
33
Petabyte Flash Capacity
Petabyte/sec IO
Fastest Storage System
on the planet
CM0
CM1
CM
CM
SM1A
SM1B
SM0A
SM0B
0 17 18 35
CM0
CM1
CM
CM
SM1A
SM1B
SM0A
SM0B
0 17 18 35
CM0
CM1
CM
CM
SM1A
SM1B
SM0A
SM0B
0 17 18 35
CM0
CM1
CM
CM
SM1A
SM1B
SM0A
SM0B
0 17 18 35
CM0
CM1
CM
CM
SM1A
SM1B
SM0A
SM0B
0 17 18 35
CM0
CM1
CM
CM
SM1A
SM1B
SM0A
SM0B
0 17 18 35
CM0
CM1
CM
CM
SM1A
SM1B
SM0A
SM0B
0 17 18 35
CM0
CM1
CM
CM
SM1A
SM1B
SM0A
SM0B
0 17 18 35
CM0
CM1
CM
CM
SM1A
SM1B
SM0A
SM0B
0 17 18 35
NSF/TACC DELL/DSSD Storage Cluster
Direct Support for Hadoop/HDFS
Copyright 2016 DSSD+EMC, Inc. Proprietary/Confidential: NDA Use Only
HDFS Architecture
10
HDFS
Filesystem
Buffer Cache
Device Driver
SATA Controller
Disk
HDFS
Filesystem
Buffer Cache
Device Driver
HDFS
PCIe SSD
PCIe
SATA
PCIe
PCIe
Hadoop
Kernel
Motherboard
Traditional PCIe SSD DSSD✓✘ ✘
DSSD HDFS Plugin delivers 2X faster I/O to the application
Application Level Benefits
• 10X Improvements in HDFS Performance
• 3X Improvement in Oracle Benchmarks
• Consistent 100 usec Latency to Application
• Shared Data Avoids Need for Replication
• Enterprise-grade Data Integrity and Security
Use Cases
Oil & Gas: Extraction Simulation Grids
Financial: Transaction Modeling
Government: Design
Life Sciences & Research
Fraud Detection Risk Analytics Predictive ModelingReal Time Analytics on Streaming Data
Financial Modeling
Ingest Market Data
Analyze in near real-time
Correlate with Historical Data
Order of Magnitude Speed-upover existing solutions
Customer Quotes
“Cloudera has tested the DSSD D5 appliance in our lab, and we've seen an order of magnitude increase in performance. It’s the fastest HBase cluster we’ve ever tested.” Mike Olson, Chief Strategy Officer at Cloudera !“DSSD D5 has fundamentally changed our business by eliminating unnecessary software, hardware and pre-processed batch jobs. With DSSD, we’re able to support applications and analytics at never-before-seen speeds.” Brian Dougherty, Chief Technical Architect, CMA !"No other supplier has a storage array equivalent in performance to the D5; it stands alone." Chris Mellor, The Register
Summary: Technology Drivers
Bigger Flash Memory
Faster CPUs
Faster Switch Chips
Denser Memory
Data Centric System Architecture
Big DataCPU
PBs
TBsGBs
Memory
DiskTape
PetaByte High-Perf. Storage
CPU
CPUCPU
CPU CPU
CPU
CPU
CPU
Combination of NVMe Flash, 3D XPoint,High-performance Interconnect and new
Software APIs solve the Data Deluge
Narrow the Data Analysis Gap
Time
Data being collectedThe G
ap
Data that can be analyzed
Conclusions
1. Rapid Growth in Real-time and Historical Data!
2. Many Opportunities to Monetize such Data!
3. Reducing Latency to Applications is Key!
4. Rack-scale Storage provides 10X Speedup !
5. Expect Further Performance Improvements