A webinar hosted by MemVerge, Intel, NVIDIA, and The Next Platform. Timothy Prickett Morgan, co-editor of The Next Platform, provides his view of the Big Memory category. Mark DeMarseilles of Intel gives an update covering new Optane Persistent Memory Series 200. Rob Davis of NVIDIA explains why Big Memory needs low latency networks to distribute messages, to replicate data, and for high-availability, all without jitter. The Charles Fan of MemVerge describes Memory Machine software and different use cases including faster crash recovery, higher VM density, and high-frequency trading.
SpotFlow: Tracking Method Calls and States at Runtime
Tech Talk: Moneyball - Hitting real-time apps out of the park with Big Memory
1. 2020 Opening Day
Tech Talk
2020 MLB Opening Day
Tech Talk
Moneyball
Hitting real-time apps out of the
park with Big Memory
2. 2020 Opening Day
Tech Talk
Official Batting Order
Joe Barnes
Director
MemVerge
Mark DeMarseilles
Technical Sales
Intel
Charles Fan
CEO
MemVerge
Rob Davis
VP
Mellanox
Timothy Pricket Morgan
Co-Editor
The Next Platform
Manager Leading off
with tech
author PoV
Swinging for
the fence
with new
Optane 200
Batting clean-up
with Big Memory
software & 3 use
cases
Base stealer
with low-latency
networking for
Big Memory
3. 2020 Opening Day
Tech Talk
Have Some Fun on Opening Day
What happens in memory
stays in memory
Attend the entire Tech Talk
and you get a MemVerge
“What happens in memory
stays in memory” t-shirt.
Every attendee is automatically
registered for the raffle. If you stay
for the entire Tech Talk you might
win an MLB jersey of your choice.
Fill out your score, batting and
pitching predictions in our opening
day poll. The person or people that
best match the actual opening results
win an MLB jersey of their choice.
4. 2020 Opening Day
Tech Talk
https://bit.ly/openingdaytechtalk
❑ The starting pitcher on the winning team will pitch this many innings:
❑ The combined total number of runs scored by both teams will be:
❑ This many home runs will be hit during the game:
❑ There will be this many double-plays executed during the game:
❑ The combined number of hits by both teams will be:
Opening Day Predictions
Most right answers wins an MLB jersey of your choice
choice
5. 2020 Opening Day
Tech Talk
Big Memory
Timothy Prickett Morgan
The Next Platform
2020 MLB Opening Day
Tech Talk
7. 2020 Opening Day
Tech Talk
Driven by the Growth of Real-Time Data
IDC: Digital Transformation Driving New “Big Memory” Requirements
0%
5%
10%
15%
20%
25%
30%
0
5,000,000
10,000,000
15,000,000
20,000,000
25,000,000
30,000,000
35,000,000
40,000,000
2015 2016 2017 2018 2019 2020 2021 2022 2023 2024
Shareofreal-timedata(%)
Real-timedata(PB)
Worldwide Real-Time Data and Share, 2015-2024
Real-time data (PB) Share of real-time data with Global Datashphere (%)
• In 2024 there will be 143
zettabytes of data
created. Worldwide, data
is growing at a 26.0%
CAGR
• In 2024, share of real-time
data grows to almost 25%
• By 2021, 60-70% of the
Global 2000 will have at
least one mission-critical
real-time workload
8. 2020 Opening Day
Tech Talk
Persistent Memory Revenue on the
Rise
$0.00
$500.00
$1,000.00
$1,500.00
$2,000.00
$2,500.00
$3,000.00
Revneue ($M)
2019 2020 2021 2022 2023
REVENUE($M)
$2609M
248% CAGR 2019-2023
$65M
IDC: Digital Transformation Driving New “Big Memory” Requirements
9. 2020 Opening Day
Tech Talk
Long Term: PMEM = DRAM
Shipments
Emerging Memories Find Their Direction: Objective Analysis and Coughlin Associates
Emerging memories are well
on their way to reach $36
billion of combined revenues
by 2030.
3D XPoint memory’s sub-
DRAM prices are expected to
drive revenues to over $25
billion by 2030
PMEM
DRAM
11. 2020 Opening Day
Tech Talk
Storage
Memory
Persistent MemoryImproving
memory capacity
DRAM
HOT TIER
HDD / TAPE
COLD TIER
SSD
WARM TIER
Delivering
efficient
storage
Intel® 3D Nand SSD
Improving
SSD performance
1 Basedon Intelinternalestimates
Re-architecting the memory/storage hierarchy
11
1 Basedon Intelinternalestimates
Bottlenecks to MEMORY
CONSTRAINED
WORKLOADS demands
new Memory tier
12. 2020 Opening Day
Tech Talk
Introducing Intel® Optane™ Persistent Memory
200 Series
12
1.MT/S is Million Transfers / Second. For Cedar Island, 2933 2DPC & 3200 1DPC is POR. Stretch target is 3200 2DPC
*Many applications w illbenefit fromhigher memory frequency and bandwidth, those w ith burstyworkloads can see an additional 20%-30% bandw idth increase using Intel Bandw idth
boost, Applications w hose OEMS implement eADR can see further application performance improvement
Intel® Optane™ persistent memory 200 Series
Up TO 3200 MT/S
DDR4
12-15 wattsThermal Design Power
Up to 38% Bandwidth *Improv ement vs Intel ®Optane ™ persistent memory
100 series
Enhanced performance
features
(Improved APP Performance)
Faster
Higher Memory Frequency)
More energy efficien
(Optimized Thermals)
128GB
256GB
512GB
Modules
DDR4 Slots
14. 2020 Opening Day
Tech Talk
Multiple Operating Modes To Best Fit your Needs
14
MEMORY MODE
PLATFORM/OS/APP ACCESS
TO
High SPEED, HIGH CAPACITY
MEMORY
High capacity
Targeting >1.2X More VMs1
Affordablecapacity
128GB, 256GB and 512GB
modules
Ease of adoption
No code changes required
1Performance results are based on testing as of dates show n in configuration and may not reflect all publicly available security updates. See configuration disclosure for details. No
product or component can be absolutely secure. For more complete information about performance and benchmark results, visit w ww.intel.com/benchmarks.
“APP DIRECT”
MODE
SW enabled dIRECT ACCESS
TO data stored persistently
Persistenceprovides
high availability/
less downtime
Significantly faster
storage
Application Optimized
Direct access to higher
data capacity
15. 2020 Opening Day
Tech Talk
Operating
Systems
Databa
ses
Infrastructure &
Storage
AI / Analytics
Software Ecosystem for Intel® Optane™
Persistent Memory
2
*Softw are support may vary per platform
16. 2020 Opening Day
Tech Talk
16
Supported on future
Intel® Xeon® Scalable Processors
Platinum and Gold SKUs
intel.com/optanedcpersist
entmemory
17. Rob Davis – VP Storage Technology, Nvidia Networking
BIG MEMORY MLB
OPENING DAY TECH TALK
18. 18
BIG MEMORY NEEDS TO BE ABLE TO SCALE
Otherwise Its Amazing Performance is Limited to Small Applications
▪Scale-out
▪Capacity expansion focused on adding new
hardware instead of increasing existing
hardware capacity
▪MemVerge application provides sharing
functions
Deep Bench
Cluster of shared Big Memory
19. 19
BIG MEMORY NEEDS TO BE TIERABLE
Otherwise Its Amazing Performance is Limited to Smaller Data Sets
▪Tiering
▪Efficiently using memory
and storage technologies
according to importance
or use frequency of the
data
▪MemVerge software
provides Tiering functions
Nolan Ryan FastBig Memory
Performance
20. 20
BIG MEMORY NEEDS HIGH AVAILABILITY
Otherwise Its Persistence is Not Usable in the Enterprise
▪High Availability
(HA)
▪Storage system that is
continuously
operational or has very
little downtime
▪Redundancy is a key
feature of HA Big
Memory
▪Allows data to be kept
in more than one place
and eliminates a single
points of failure
▪MemVerge software
provides HA function
Backup Pitcher
Big Mem.
21. 21
CORE REQUIREMENTS FOR NETWORKING BIG MEMORY
▪Big Memory is really fast
▪Needs ultra low-latency network
▪Needs ultra high bandwidth network
▪Needs predictability, even under load
▪Needs Fairness (QOS)
▪Needs Very Low Packet Loss
Base Stealing Speed
Big
Memory
(PM)
22. 22
FASTER NETWORK HARDWARE SOLVES HALF OF THE BIG
MEMORY NETWORK PROBLEM…
Ethernet & InfiniBand
End-to-End 25, 40, 50, 56, 100, 200Gb
Going to 400Gb next year
27. 27NVIDIA CONFIDENTIAL. DO NOT DISTRIBUTE.
RDMA NOW COMMON ACROSS ALL STORAGE TYPES
Big Memory
(PM)
RPM
Block
File
Object
RDMA
SMB (CIFS)
NFS
Ceph
iSERSMB Direct
Ceph
over
RDMA
NVMe-
oF/RDMA
Swift
▪ RDMA for optimal
performance
• InfiniBand & RoCE
o NVMe-oF
o Nvidia GPU Direct Storage
▪ Persistent Big Memory
(3D-XPoint)
S3
FC
iSCSI
NFSoRDMANFSoRDMA NVMe-
oF/TCP
28. 4 Reasons Why Big Memory Needs RDMA
1. Pub/Sub
Distributing messages
to many servers with
low-latency and jitter
3. HA
Replicating memory to
other servers for failover
4. Scale-Out
Create memory lakes by
pooling server memory
RDMA
RDMA
2. App Migration
Memory Machine migrates app
+ memory to other servers
RDMA
RDMA
29. 29
ZERO TOUCH ROCE MEANS NO SWITCH CHANGES NEEDED
ConnectX-5, All-to-all, 8 nodes, provides results similar to lossless
30. 30
IMPORTANCE OF NETWORK LATENCY WHEN NETWORKING
BIG MEMORY
Logarithmicscale
Nvidia Ethernet Switch & Adapter
Network hops multiply latency
Request/Response
Big Mem.
200Gb
100Gb
Run Down
InfiniBand900ns
1.3μs
Common Ethernet Switch & Adapter
32. 2020 Opening Day
Tech Talk
Big Memory Software
& Use Cases
Charles Fan
MemVerge
2020 MLB Opening Day
Tech Talk
33. 33
A software-defined
memory service that
virtualizes DRAM and
Persistent Memory
• Compatible with DRAM,
no application rewrite
necessary
• Larger capacity, Similar
Performance and Lower
Cost than DRAM
Featuring Data
Services designed
for memory
• World’s first ZeroIO™
in-memory snapshot
• Fast crash recovery to
any snapshot
• Clone new application
instance from any
snapshot without
replicating memory
MemVerge Memory Machine™
34. 2020 Opening Day
Tech Talk
Uses Cases: Real-Time Applications
AI/ML inference like fraud
detection,image
classificationand
recommendationengines
Latency-sensitive
transactional workloads such
as trading applications in
capital markets
According to IDC, by 2021, 60-70% of the Global 2000 organizations
will have at least one mission-critical real-time workloads like these.
Memory-intensive
applications in media &
entertainment, healthcare,
and oil & gas
35. 2020 Opening Day
Tech Talk
The Need
• Limit disruption to business
when an in-memory app
crashes
The Problem
• It takes a long time for an in-
memory app to restart after
crash or planned shutdown
Memory Machine Use Case
Crash Recovery
36. 2020 Opening Day
Tech Talk
MemVergeSolution
• AutoSave: Memory Machine takes instant
ZeroIO™ snapshots as frequently as every
1 minute
• Application is automatically restarted from
the last snapshot after crash, and it is
FAST!
Memory Machine Use Case
Crash Recovery
1TB
39. 2020 Opening Day
Tech Talk
The Need
• To create additional
instances of In-Memory
Databases for other
applications, DevOps,
QA, etc.
Memory Machine Use Case
Cloning
• Multipleinstances
of physical memory
are expensive
• Cloning databases
is time-consuming
The Problem
40. 2020 Opening Day
Tech Talk
The MemVerge Solution
• ZeroIO™ snapshots to
memory as frequent as
every minute
• A new in-memory database
instance can be created
from any past snapshot
without a physical memory
copy
Memory
Snapshot
+
Clone
Replay
Restart
Memory Machine Use Case
Cloning
41. 2020 Opening Day
Tech Talk
Cloning a Redis Database with 300M Keys
41
900,240 (15 min)
581 (0.5 sec)
0
100,000
200,000
300,000
400,000
500,000
600,000
700,000
800,000
900,000
1,000,000
DRAM PMEM (Memory Machine)
Milliseconds
1,500X Faster
No duplication
of memory
resources!
42. 2020 Opening Day
Tech Talk
The Need
• Optimize server utilization by increasing
VM density & decreasing cost per VM
The Problem
• Memory capacity is the bottleneck
Memory Machine Use Case
KVM
43. 2020 Opening Day
Tech Talk
The MemVerge Solution
• Memory Machine™ software multiplies
the amount of memory available to the
hypervisor
• Adjustable DRAM:PMEM ratio enables
memory performance similar to DRAM
Memory Machine Use Case
KVM
DRAM PMEM
Memory Machine
44. 2020 Opening Day
Tech Talk
MySQL performance
44
0
20000
40000
60000
80000
100000
120000
140000
160000
10G Data 40G Data
SysbenchQPS
DRAM PMEM Cache-2G Cache-4G Cache-8G Cache-16G
45. 45
Memory Machine™
Enterprise-Class Memory Virtualization Software Key Benefits to Real-TimeApps
Scalability
Terabytes of DRAM-speed memory
available to memory-intensive apps
Availability
ZeroIO™ Snapshot enables faster crash
recovery
Agility
Instant Clones without duplicating of
physical memory
Compatibility
No changes to application necessary
46. MemVerge Vision for Big Memory Industry
Compute
Memory
Performance Storage
Capacity Storage
Compute
Big Memory
Capacity Storage
48. 2020 Opening Day
Tech Talk
Thank-you
Try it!
Contact Andrew.Degnan@memverge.com to sign-up for a PoC
49. 2020 Opening Day
Tech Talk
Learn More About Big Memory
49
Persistent Memory
Breakthrough Memory Optimized for Data-Centric Workloads
https://www.intel.com/content/www/us/en/architecture-and-technolog y/optane- dc-persistent-memor y.html
Intel® Optane™ PersistentMemory Workload Solutions
https://www.intel.com/content/www/us/en/architecture-and-technolog y/optane- persistent-memor y-soluti ons.html
Gartner's Donald Feinberg on PersistentMemoryTechnology
https://youtu.be/9oQ78M51RfQ
RDMA Networks
Improve Data Transfer Efficiency Through RDMA
https://www.mellanox.com/products/adapter-ethernet-SW/RDMA-RoCE-Ether net-Network-Efficiency
Learn how RDMA Accelerates Access to Data
https://www.mellanox.com/sites/default/files/related- docs/technical-briefs/TB_RDMA.pdf
Video – RDMA over Converged Ethernet (RoCE)
https://youtu.be/mu2v95feops
Big Memory Software
MemVerge Corporate Brochure
https://www.memverge.com/web/wp-content/uploads/2020/07/C orporate-Brochure_2020.pdf
Memory Machine Software Data Sheet
https://www.memverge.com/web/wp-content/uploads/2020/06/Data-Sheet_M emory-Machine.pdf
IDC Big Memory Definition and PMEM Forecast
https://www.slideshare.net/MemVerge/
The Next Platform Coverage
Those Without PersistentMemoryare Fated to Repeatit
https://www.nextplatfor m.com/2019/05/20/those- without- persistent-memor y-are-fated-to-repeat-it/
Getting Around The Limits ofMemory to Accelerate Applications
https://www.nextplatfor m.com/2019/08/18/getting-around-the-limits- of-memory-to-accelerate-applications/