1. Accelerating Applications with NVM Express™ Computational Storage 2019 NVMe™ Annual Members Meeting and Developer Day March 19, Prepared by Stephen Bates, CTO, Eideticom & Richard Mataya, Co-Founder & EVP, NGD Systems

3. Agenda
What??
Why??
Who??
How??

4. WHAT??

5. NVMe™ is a transport
Michael Corwell, GM Storage, Microsoft Azure, Dec 5th 2018

6. One Driver to Rule Them All?!
NVMe™ has been incredibly successful as a storage protocol.
Also being used for networking (NVMe-oF™ and things like AWS Nitro and Mellanox’s Sexy NVMe Accelerator Platform (SNAP)).
Why not extend NVMe to compute and make it the one driver to rule them all?

7. What is Computational Storage? SNIA has Defined the Following
Computational Storage Drive (CSD): A component that provides persistent data storage and computational services
Computational Storage Processor (CSP): A component that provides computational services to a storage system without providing persistent storage
Computational Storage Array (CSA): A collection of computational storage drives, computational storage processors and/or storage devices, combined with a body of control software

8. WHY??

9. Real question is “Why not NVMe?”
Why NVMe™?
Accelerators require:
Low latency
High throughput
Low CPU overhead
Multicore awareness
Management at scale
QoS awareness
NVMe provides:
Low latency
High throughput
Low CPU overhead
Multicore awareness
Management at scale
QoS awareness
Real question is “Why not NVMe?”

10. Let’s Go Fishing for Data

11. NVMe™ Computational Storage
CPU
NVMe based Computational Storage Processor (CSP) advertises zlib compression.
Operating System detects the presence of the NVMe CSP
Used by the device-mapper to offload zlib compression to NoLoad.
This can be combined with p2pdma to further offload IO.
With standardization this can be vendor- neutral and upstreamed.
DRAM
PCIe Subsystem
. . .
CMB
NVMe CSP
NVMe SSDs

12. NVMe-oF™ Computational Storage
An NVMe™ CSP is represented as an NVMe Computation Namespace. Therefore it can be exposed over Fabrics.
Compute nodes can borrow CSPs, CSDs and standard NVMe SSDs via fabrics from Computational Storage Arrays (CSAs).
NVMe Computational Storage can use the same fabrics commands that are used by legacy NVMe-oF.
Application code is identical regardless Computation is local (PCIe) or remote (Fabrics)
Ethernet TOR Switch
Compute Node
Compute Node
Compute Node
Computational Storage Array
NVMe CSPs, CSDs and SSDs

13. Example of a Hadoop Cluster - In-Situ Processing
Ability to Migrate Data Nodes into drives
Allow for user to reduce CPU Core count
Current example:

14. Network Monitoring Offload

15. WHO?

16. SNIA Computational Storage TWG

17. HOW?

18. NVMe™ for Computation: Software
Management
nvme-cli nvme-of
Applications
Userspace
libcsnvme
SPDK OS
Hardware
NVMe CSPs, CSDs and CSAs 18

19. NVMe™ for Computation: Standards
NVMe Computation Namespaces: A new namespace type with its own namespace ID, command set and admin commands. Operating Systems can treat these namespaces different to storage namespaces.
Fixed Purpose Computation: Some computation can be defined a way that an Operating System can consume it directly (e.g. zlib compression tied into the crypto API in Linux).
General Purpose Computation: Some Computation Namespaces will be flexible and can be programmed and used in user-space (/dev/nvmeXcsY anyone?)
NVMe Computation over Fabrics: User-space does not know or care if /dev/nvmeXcsY is local (PCIe) or remote (Fabrics)

20. There are many paths to Computational Storage

21. Processor Path in an NGD Systems NVMe™ SSD
It’s an NVMe SSD at the core
No impact on host read/write
No impact on NVMe driver
Standard protocols
But then there is MORE (Patented IP)
Dedicated compute resources
HW acceleration for data analytics
Seamless programming model
Scalable

22. + computation = awesome
Call to Arms!
If this all sounds interesting, please join the SNIA Computational Storage TWG.
End-users and software people are needed!
If you have thoughts on how you would consume NVMe™ Computation, please let us know
As SNIA starts interfacing with NVMe please participate in the TPAR/TP discussions! +
computation =
awesome

23. Questions?