Upload
insidehpc
View
322
Download
3
Embed Size (px)
Citation preview
©2016 Micron Technology, Inc. All rights reserved. Information, products, and/or specifications are subject to change without notice. All information is provided on an “AS IS” basis without warranties of any kind. Statements regarding products, including regarding their features, availability, functionality, or compatibility, are provided for informational purposes only and do not modify the warranty, if any, applicable to any product. Drawings may not be to scale. Micron, the Micron logo, and all other Micron trademarks are the property of Micron Technology, Inc. All other trademarks are the property of their respective owners.
Memory and Storage-Side ProcessingHow persistent memory will bring an entirely new structure to large data computingSteve Pawlowski, VP of Advanced Memory Systems
Persistent Memory Today
© 2016 Micron Technology, Inc. |
Closing the Memory / Storage Latency Gap
3
Nanoseconds Microseconds Milliseconds
Storage I/O Memory Persistent Memory
Memory Control(Load/Store)
I/O Control(Read/Write)
Hard Disk Drive(Spinning Media)
SATA/SAS SSD(NAND)
NVME SSD(NAND)
QuantX™ Technology
SSD(3D XPoint™
memory)
DIMM(3D XPoint™
memory)ASP/Bit
Latency(not to scale)
NVDIMM-N(DRAM/NAND)
RDIMM/LRDIMM(DRAM)
Cost Optimized PM
Performance OptimizedPM
May 1, 2023
© 2016 Micron Technology, Inc. |4 | Micron Confidential
Impact of NVDIMM-N on Application Performance
| May 1, 2023
• Data committed to persistent media written to NAND through the I/O stack
• Data committed to persistent media written to DRAM on NVDIMM-N through the I/O stack
• Latency improvement due to writing to DRAM (35ns) versus NAND (85,000ns)
• Data committed to persistent media written to DRAM on NVDIMM-N through Load/Store Bus
• Latency improvement due to circumventing overhead associated with I/O stack
35ns
35ns
85us
85us
35ns
DRAM + NAND-Based NVMe SSD
DRAM + NVDIMM-N Block Mode
DRAM + NVDIMM-N Direct Mode
35ns
35ns
25us
25us
35ns
35ns
35ns
35ns
35ns
35ns NV Direct
NV Block
NVMe
LATENCY
© 2016 Micron Technology, Inc. |
HPE and Microsoft Enterprise Performance
| May 1, 20235 | Micron Confidential
Hardware Broadwell 2-way Platform DB data on 6x 400GB SSD
8GB NVDIMM Cache + SSDsMirrored Pool of
SSDs
SQL Server Transaction Log
970k Tx/min372us write latency
1.08M Tx/min181us write latency
+11% Performance
Benefit(Block Mode, Read Centric)
Exchange Server
+63% Performance
Benefit(Block Mode, Write Centric)
780k Tx/min 1.3M Tx/min
Future Vision
© 2016 Micron Technology, Inc. |
Fast Persistent Writes Metadata StorageWrite Back Cache
Scale-out Storage
Big Data Analytics
In-Memory Databases
Relational Databases
Persistent Memory will … Persist
Higher capacities and storage-class memories will open even broader applications
BECOMING THE MAINSTAY OF MEMORY SUBSYSTEMS
© 2016 Micron Technology, Inc. |May 1, 2023 | Micron Confidential
The Critical Element of Future Memory Systems
8
NVDIMM-NHybrid Memory Cube
This opens interesting new possibilities for managing data
Integrated on-board controller
© 2016 Micron Technology, Inc. |May 1, 2023 | Micron Confidential
Moving Data to the Processor is Costly
What if we could avoid this 1000x chasm?
9
One floating-point calculation
17 picojoules
Moving data from DRAM to CPU
17,000 picojoules
© 2016 Micron Technology, Inc. |May 1, 2023 | Micron Confidential
Why Hasn’t it Been Done Before?
DRAM transistors are not ideal for logic processing– In-memory processing must be very simple– Processing must be intrinsically data-centric– Complementary to CPU / Custom Processors
10
Memory Technology
Low leakage transistorsLow power
Processor Technology
High speedHigh power
© 2016 Micron Technology, Inc. |May 1, 2023 | Micron Confidential
What Can Be Done with STM?
Data-centric problems• Data sort• Databases• Machine learning (90% of the problem is feeding the MACs)
Computationally lightweight problems Solutions that can be smoothly transitioned:• Replace a library call (qsort ->msort)• Replace a subsystem (MySQL)
11
© 2016 Micron Technology, Inc. |May 1, 2023 | Micron Confidential
An Illustration
12
Challenges to Overcome
© 2016 Micron Technology, Inc. | | Micron Confidential
Driving Easier Market Adoption
– Simplify the power source
– Minimize SW and programming ecosystem impact
– Appropriate energy/performance tradeoffs
– Latency optimization
| May 1, 202314
HPE Smart Storage NVDIMM Integrated Backup Power
© 2016 Micron Technology, Inc. |May 1, 2023 | Micron Confidential
Implications of NVDIMM-P
High capacities enable truly large in-memory computing Latencies introduce opportunity for tiered memory systems Industry ecosystem should begin preparing
today
15