adopting persistent memory in new memory-converged … · 2019. 12. 21. · • by 2025, storage...
TRANSCRIPT
2019 Storage Developer Conference. © Insert Your Company Name. All Rights Reserved. 1
Adopting Persistent Memory in New Memory-Converged Infrastructures
Charles FanMemVerge
2019 Storage Developer Conference. © Insert Your Company Name. All Rights Reserved. 2
Data Infrastructure Pain Points
DRAM is small and expensive(100’s of GBs)
Storage IO is slow(100’s of microseconds)
Could there be a solution that makes storage faster and memory bigger?
Machine Learning, Big Data and IoT demand nanosecond speed + petabyte scale data infrastructure
2019 Storage Developer Conference. © Insert Your Company Name. All Rights Reserved. 33
Storage Class Memory has Emerged!
• Intel Delivered Optane DC Persistent Memory based on
3D XPoint tech in Q2 2019.
– Revenue projected to reach $3.6B by 2023
• Additional major vendors to join the foray by 2022
– Potentially a $10B+ market by 2025
• Software ecosystem will be key for technology adoption
– Will disrupt software stacks
– How to adopt without requiring application rewrite?
2019 Storage Developer Conference. © Insert Your Company Name. All Rights Reserved. 4
MemVerge software leverages Storage Class Memory technology to deliver larger memory and faster storage to
applications without requiring application rewrites
World’s FirstMemory Converged Infrastructure
Memory “Hypervisor”
SCM-native Distributed File System
SCM-native Distributed Memory System
2019 Storage Developer Conference. © Insert Your Company Name. All Rights Reserved. 5
Memory Converged Infrastructure PlatformSearch and QueriesMachine
LearningBig Data Analytics
...
…
Distributed Memory Objects (DMO)
DRAM
Compute Node 3
SCMSSD
DRAM
Compute Node 4
SCMSSD
DRAM
Compute Node 5
SCMSSD
DRAM
Compute Node 2
SCMSSD
DRAM
Compute Node 1
SCMSSD
Memory “Hypervisor”
SCM-native Distributed File System
SCM-native Distributed Memory System
2019 Storage Developer Conference. © Insert Your Company Name. All Rights Reserved. 6
AI Training with Checkpointing
Problem Model training takes a long time to complete for
large datasets Failure recovery is painful without frequent
checkpointing Data preprocessing and importing can take a long
time Delayed model deployment
Solution MemVerge DMO, powered by Optane DC
persistent memory, improves checkpointing speed and data loading speed.
up to 6XTraining Speed
InstantCheckpoint Recovery
up to 350XData Import Speed
2019 Storage Developer Conference. © Insert Your Company Name. All Rights Reserved. 7
Problem Spark SQL Out of DRAM Disk I/O too slow Data spill degrades performance Local SSDs wear out by frequent intermediate data
writes
Solution Adding MemVerge DMO to the Spark cluster
accelerates the entire cluster Moving intermediate state off Spark Elastic Computing
nodes increased the cloud elasticity of the solution.
Big Data Analytics with Spark
5XTerasort Speed
7XRDD Caching Speed
100%Cloud Elasticity
2019 Storage Developer Conference. © Insert Your Company Name. All Rights Reserved. 8
The MCI Vision• By 2025, Storage Class Memory will be mainstream. Data Infrastructure will be memory-centric.
• Performance-tier storage will be replaced by Memory Converged Infrastructure (MCI) co-located with compute. Memory capacity will be expanded by the same MCI layer.
• MemVerge aspires to be a leader in MCI.
ComputeMemory
Performance-tier Storage
Capacity-tier Storage
ComputeMCI
Capacity-tier Storage
Today Tomorrow
2019 Storage Developer Conference. © Insert Your Company Name. All Rights Reserved. 9
MCI Starts Today
as low as 1 µs Access Latency
up to 768TB Total Cluster Memory
up to 10M+ IOPS Per Node