welcome to the sssi pm & nvmdimm sig open call we will ......welcome to the sssi pm &...
TRANSCRIPT
Welcome to the SSSI PM & NVMDIMM SIG
Open CallWe will begin shortly
Computational StoragePart of the Evolving Ecosystem
10 May 19
Scott Shadley, Co-ChairNick Adams, Co-Chair
© 2019 Storage Networking Industry Association. All Rights Reserved.
Agenda
Introduction to Computational StorageHow long has this idea been around, Why Now?How the TWG was formedWhat is the link to PM/NVDIMM?
Computational Storage Working Group FocusTaxonomy is Key, need the right TLAsScope is Critical, Roadmap to successA look at architectures – including PM/NVDIMM Synergies
Solutions Today and PlannedA look at some current solutionsA view of people’s thoughts of future solutions
3
Why Do We Need Computational Storage?The Back Story and Introduction
© 2019 Storage Networking Industry Association. All Rights Reserved.
Many Factors driving a Need for Computational Storage
5
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 )
© 2019 Storage Networking Industry Association. All Rights Reserved.
What is Computational Storage
When to USE Computational StorageLarge Data Transfers, PCIe is bottleneckData pre/post processing & analysis Data can bypass the host - video deliveryAbility to move Software App to Storage
When to USE NVDIMMCompute heavy with small data-transferSmall data compute - in-memory compute Little to no parallelism
6
© 2019 Storage Networking Industry Association. All Rights Reserved.
The Evolution of Computational Storage
A delicate process to build an Ecosystem
Great ideas! Time was needed to build it Many technology papers exist around:
“Active Disks”, “CAFS”, “Near-Data”“In-Storage”, “In-Situ”, “Near-Storage”
So did some initial products!
7
© 2019 Storage Networking Industry Association. All Rights Reserved.
How to Drive the Real Market
So what do we do now? The answer came through conversations at many companies…
“Hey aren’t we all working to the same goal?” Let’s work together to go forward!
And BoF… @ 7AM… DURING FMS was proposed8
© 2019 Storage Networking Industry Association. All Rights Reserved.
Time for Impactful Growth
A packed room got together & agreed: Move forward, fast!Equal parts start-up, and mature HWThis group asked questions like? “Could we create a
design to solve problems by bringing the compute into storage?
“Could this greatly reduce I/O bus traffic & allow distributed processing?”
9
© 2019 Storage Networking Industry Association. All Rights Reserved.
SNIA Gets Formally Involved
Multiple Computational Storage sessions
And another BoF, this time at 7PM, even more attended!
Launched at the SNIA SDC18 Event as the proper proving ground for the development of standardization work
10
© 2019 Storage Networking Industry Association. All Rights Reserved.
Playing Nice Together is Needed!
Is this a solution replacing a solution?
Complimentary work to the pyramid
Another facet of advancement of compute
In-Memory is needed, but some work canbe offloaded all the way to storage!
11
Com
puta
tiona
l St
orag
e
So Now What?The Progression of the TWG
© 2019 Storage Networking Industry Association. All Rights Reserved.
Compute, Meet Data
Based on the premise that storage capacity is growing, but storage architecture has remained mostly unchanged dating back to pre-tape and floppy…
How would you define changes to take advantage Compute at Data?
13
© 2019 Storage Networking Industry Association. All Rights Reserved.
Let’s Go Fishing for Data
14
© 2019 Storage Networking Industry Association. All Rights Reserved.
Now What, How do we Begin?
Born Oct 2018, the Computational Storage Technical Working group had 10 Founding companies.
Interest and attention grew… Quickly…
Meetings moved to weekly (AM/PM)
Now the single largest SNIA TWG!!15
© 2019 Storage Networking Industry Association. All Rights Reserved.
40+ Participating Companies128+ Individual Members
16
© 2019 Storage Networking Industry Association. All Rights Reserved.
Finding a Focus and Direction
Focus on a definition list to ensure we cover questions from users on what it is and what products there are
Drive to a Scope and path to universal usage modelToday we have custom… Tomorrow Standard… Sound Familiar
17
© 2019 Storage Networking Industry Association. All Rights Reserved.
A New Taxonomy is Needed
Computational Storage:Architectures that provide Computational Storage Services coupled to storage offloading host processing or reducing data movement.
Computational Storage Service (CSS)Fixed (FPCSS)General Purpose (GPCSS)
18
© 2019 Storage Networking Industry Association. All Rights Reserved.
A New Product Category
Computational Storage Device (CSx)
Computational Storage Drive (CSD)Computational Storage Processor (CSP)Computational Storage Array (CSA)
19
Co m pu t a t ional S t o r age Arr a y(Acc e ss v ia C S P and / o r di r e c t t o S t o r age )
O p t ional Pr o x iedS t o r age Acc e s s
O p t ional Di r e c tS t o r age Acc e s s
Co m pu t a t ional S t o r age D r i v e(Acc e ss v ia C S P and / o r di r e c t t o S t o r age )
Co m pu t a t ionalS t o r age Pr o c e ss o r
Co m pu t a t ionalS t o r age D r i v e
T r adi t iona lS t o r age De v i c e
Ho s t A gen t 1
F ab r i c (P C I e , E t he r ne t, e t c )
Ho s t A gen t 2 Ho s t A gen t N
S t o r ag e
I/ O I/ OM G N T M G N T I/ OM G N T
O p t ional Pr o x iedS t o r age Acc e s s
O p t ional Di r e c tS t o r age Acc e s s
I/ OM G N T I/ OM G N T
S t o r ag e
I/ OM G N T I/ OM G N T
Arra yCont r o l
C S
C S
C S
C S
C S
C S
C S
C S
Co m pu t a t iona lS t o r ag e
Pr o c e ss o r(s )
Co m pu t a t iona lS t o r ag e
Pr o c e ss o r(s )
Co m pu t a t iona lS t o r ag e
Pr o c e ss o r(s )
Co m pu t a t iona lS t o r ag e
Pr o c e ss o r(s )
C SC SD r i v e r s C S
D r i v e r s C SD r i v e r
A nd /O r A nd /O r
© 2019 Storage Networking Industry Association. All Rights Reserved.
Scoping What Can Be Done
Multiple F2F sessions have been focused on what we can accomplish and what we will leave for later
ManagementSecurity Operation
20
© 2019 Storage Networking Industry Association. All Rights Reserved.
Architectures Aligned
Define the high level Open architecture firstAddress interactions of compatible solutions
PM Media, PM SolutionsWork with NVDIMM platform adaptations
Then spin off the different interfaces, protocols, etc. to the right industry groups, such as scala.org, DMTF, and NVMe.
21
© 2019 Storage Networking Industry Association. All Rights Reserved.
NVMe for Computation: Software
CSDs, CSPs and CSAs Hardware
OS
libcsnvmeSPDK
ApplicationsManagement
nvme-clinvme-of
Userspace
22
Solutions Today and TomorrowExamples from TWG Members
© 2019 Storage Networking Industry Association. All Rights Reserved.
Many Paths to Computational Storage
24
© 2019 Storage Networking Industry Association. All Rights Reserved.
NVMe FP-CSP RAID Offload
NVMe based Computational Storage Processor (CSP) advertises Fixed Purpose accelerator capable of RAID parity generation.Operating System detects the presence of the NVMe FP-CSPUsed by the device-mapper to offload parity calculations for RAID.This can be combined with p2pdma to further offload IO
25
CPU
PCIe Subsystem
DR
AM
. . .CMB
NVMe CSP NVMe SSDs
md-raid
© 2019 Storage Networking Industry Association. All Rights Reserved.
Efficient Utilization of Compute
Systems requiring similar sets of Computational Units can take advantage of large scale dynamically configurable CSAs utilizing FaberXTM network.Large scale CSA shared by multiple hostsHigher Compute Utilization achieved by pooling CSPs for HostsHigher Storage Utilization achieved by pooling CSDs for CSPsCSDs and CSPs ability to do p2p data exchangeLarge, dynamically configurable, variety of Fixed Purpose CSDs will be available for each Host Large Scale Dynamically
Configurable CSA
FabreXTM
DRAMDRAMDRAMDRAM
. . .
CSPs CSDs
26
© 2019 Storage Networking Industry Association. All Rights Reserved.
AI Inference at the Storage
Generate Metadata database (e.g tags) over a large set of unstructured data locally with an integrated AI inference engineOperation may be:
Triggered by a host processorDone offline as a background task (batches)
Metadata database may be then used by upper layer Big Data Analytics software for further processingCan work both on direct attached storage or on remote over the network storageExamples: Video search, Ad insertion, Voice call analysis, Images, Text scan, etc
27
Computational Storage Array
Network
CPUCSDs
.
.
.
.
© 2019 Storage Networking Industry Association. All Rights Reserved.
Video Transcoder Processor
Computational Storage Processor (CSP) designed for scalable high-efficiency Video Encoding/TranscodingIntegrated with FFmpegLinear scalable encoding capability, offload CPU from heavy encoding computingSave 90% of power consumptionsDeterministic low latency
28
PCIe Interface
Codensity™ Video TranscodersNVMe FP-CSP inside U.2 module, w/o SSD functions
Qty=1..N
© 2019 Storage Networking Industry Association. All Rights Reserved.
Real-Time AI Genomics Improvement
29
DNA and Protein alignment Database Management
Up to 100% more performance at no cost in CPU or Memory
Resolves the IO Bottleneck between CPU and Storage
The Basic Local Alignment Search Tool (BLAST).
Compute in Storageremoves CPUs lack of bandwidth to the data
© 2019 Storage Networking Industry Association. All Rights Reserved.
Hadoop: Job Throughput
30
3.1 w/ EC (6+3)
All benchmarks configurations use HDD as main storage24 Mapper/Reducers per Datanode *9 = 216 total
Better performance on CSS reported with lower Mapper/Reducers possible
Datanode Config:Dual E5-2640v3, 128GB DRAM, 12*6TB
SAS HDD
…
14% ↓ vs. baseline116% Job Throughput
23% ↓ vs. baseline131% Job Throughput
Compute Offload AND Flash Temp
37% ↓ vs. baseline160% Job Throughput
Baseline: Compute & Storage I/O
Bound Compute Offload Only
Flash Temp Only
One per server, 9
total
© 2019 Storage Networking Industry Association. All Rights Reserved.
In Summary – Call to Action
Computational Storage is a Real MarketCustomers are deploying today
Solutions exist and will continue to growMaking the ‘uniform’ helps adoption
Standardizing the host interaction is vitalWe NEED more Support from Users/SW Solutions
Working with all TWG/SIGs & initiatives is keyJoining forces and cross-membership adds to success
31
Thank You!!www.SNIA.org/Computational
www.SNIA.org/forums/sssi/NVDIMM