Challenges in high-precision communication/computing and programmability.

​

Prof. Christian Esteve Rothenberg (DCA/FEEC/UNICAMP)

Next Generation Routing and Management Infrastructures panel in 6G – NET-GENES22 Workshop

@IEEE Future Networks Word Forum 2022.

13-Oct 2022

Agenda

• What is high-precision communication/computing?
• Clear vision and definitions?
• How to realize high-precision C/C?
• From Gaps in existing technologies to Challenges
• Trade-offs in delivering high-precision C/C?
• Determinism vs. Programmability

Introduction

• New networking applications with challenging requirements
• Industrial Internet applications based in controllers with ultrafast control loops subject to rigid timing requirements.
• Mission-critical Tactile Internet applications such as tele-surgery require ultra-low latencies without any loss in connectivity.
• “Metaverse”
• ...
• "Best Effort" internetworking technology no enough
• Need for "high precision" networking services that are able to guarantee the ambitious service levels and precise timing constraints demanded by envisioned applications.
• New networking architectures and protocols?
• A new Clean Slate / Future Internet wave of R&D?
• Internet QoS revamped?
• Technological gaps under discussion, e.g. ITU-T Technical Report FG-NET2030-Gap (2020)
• Implications of high-precision networking for management, for measurements, for service assurance and fulfillment? Cf. [Clemm et al, 2020]
• Standardization: IETF Detnet & IEEE TSN

NETWORKING 2030-2040

​

​

​

Source: [METAVERSE] I.F. AKYILDIZ. “METAVERSE: Challenges for Extended Reality and Holographic-Type Communication in the Next Decade”

​

​

Challenges in CFPs around TSN/DetNet

e.g., 2ND INTERNATIONAL WORKSHOP ON TIME-SENSITIVE AND DETERMINISTIC NETWORKING (TENSOR) 2021

https://networking.ifip.org/2021/workshops/international-workshop-on-time-sensitive-and-deterministic-networking-tensor.html

• Emerging standards and advances on IEEE TSN �(e.g., 802.1Qbv, 802.1Qbu, 802.1Qci etc)
• Layer 3 Deterministic IP Networking
• TSN in Industrial Automation, Automotive, Aerospace, Service Provider Networks and TSN convergence with other technologies like Wi-Fi and 5G (Network Service KPIs translation etc.)
• Enabling 5G Time Sensitive Communications
• Hardware/Software co-Design for TSN
• Testbed for TSN and deterministic communications
• TSN Testing methodologies
• Deterministic communications for the Cloud
• Programmable dataplanes for TSN
• OPC-UA/TSN convergence
• Orchestration and Management of Software-Defined Deterministic Networks
• Highly reliable deterministic communications.
• Queueing Control, scheduling, admission control policies
• Artificial intelligence for deterministic networks
• New key techniques for time-sensitive networking

• Network Telemetry & Monitoring for IEEE TSN and/or IETF Detnet
• Ultra-reliable low-latency communications for 5G
• SDN applications for high-precision, high-performance networking
• Wireless deterministic connectivity (WiFi6, emerging WSN)
• Internetworking of IEEE TSN networks with mobile/wireless technologies
• Emerging Standards for Wired and Wireless Time Sensitive Networking (TSN)
• Security aspects of deterministic networking and real time systems
• Segment routing for Detnet
• Time-sensitive networking system implementation experience
• Time-synchronization and robustness for low-latency TSN
• Network calculus for TSN bounds

High-precision measurements

[FG-NET2030-Gap G.MGMT.1] There is gap in measurement techniques that are capable of measuring end-to-end service levels that combine all of the following.

• Accuracy: The combination of high measurement accuracy (precision on the order of 10 microseconds end-to-end) with acceptable measurement overhead;
• Coverage: lack of measurement techniques that allow for complete coverage not only of selected service instances but across the board, specifically for mission-critical, high-precision services at scale, without having to resort to sampling; and,
• Privacy-preservation: no reliance on packet snooping that could expose personally identifiable information.

Source: ITU-T Technical Report FG-NET2030-Gap (2020)

Comprehensive device telemetry at scale

[FG-NET2030-Gap G.MGMT.2] There is a requirement for techniques that allow the generation of actionable telemetry on individual devices, with the telemetry being fine-grained enough to allow for analysis of individual packet behaviour traversing the device and to able to operate at very high frequency (i.e., able to provide updates at ms or µsec frequency). Such techniques are needed to assure high-precision services.

Source: ITU-T Technical Report FG-NET2030-Gap (2020)

Comprehensive end-to-end packet telemetry at scale

[FG-NET2030-Gap G.MGMT.3] Techniques that allow the collection of telemetry data, end-to-end across network devices, for given packets, flows, and service instances, in a way that generated telemetry does not overwhelm or otherwise negatively impact production traffic, and that does allow to scale, i.e., to cover every high-precision flow and flow burst of interest.

Source: ITU-T Technical Report FG-NET2030-Gap (2020)

Inherent accounting support for SLOs

[FG-NET2030-Gap G.CAP.5]

SLO-aware accounting refers to the ability to validate whether service level objectives have been met. A separate infrastructure for accounting and for validation of adherence to SLOs is required.

High-precision services are premium services and not best-effort but "guaranteed", likely leading to the requirement to allow senders and receivers to get indications that service level objectives are being adhered to as part of the network service interface. Such facilities could include warnings (or an indication) that a specified in-time or on-time target has been compromised, or a ''receipt'' at the end of a session providing a summary of session level SLO accounting.

Source: ITU-T Technical Report FG-NET2030-Gap (2020)

From Challenges to Solutions....

​

​

​

Source: [Clemm et al, 2020b] Toward Truly Immersive Holographic-Type Communication: Challenges and Solutions

Solution Trade-offs :: There are many…

Trade-offs :: More…

Complexity

Counter-complexity

Sweet Spot

Function

Trade-offs & Universal Laws and Architectures [Turing]

Source: D. Meyer

Performance

Programmability

Trade-offs in High-Precision C/C ?

Realm of the Impossible?

Plane of the Possible?

Programmability

Cost | Scale

Precision

Source: [P3]

References

[Clemm et al, 2020a] A. Clemm, M. F. Zhani, R Boutaba: "Network Management 2030: Operations and Control of Network 2030 Services. https://doi.org/10.1007/s10922-020-09517-0 Journal of Network and System Management, Springer, March 2020.

[Clemm et al, 2020b] Alexander Clemm, Maria Torres Vega, Hemanth Kumar Ravuri, Tim Wauters, and Filip De Turck. 2020. Toward Truly Immersive Holographic-Type Communication: Challenges and Solutions. Comm. Mag. 58, 1 (January 2020), 93–99. https://doi.org/10.1109/MCOM.001.1900272

[FG-NET2030-Gap] ITU-T Technical Report FG-NET2030-Gap, 2020

[METAVERSE] I.F. AKYILDIZ. “METAVERSE: Challenges for Extended Reality and Holographic-Type Communication in the Next Decade”

[P3] Softwarized dataplanes and the P3 trade-offs: Programmability, Performance, Portability. Keynote at IEEE 18th International Conference on High Performance Switching and Routing (HPSR), Campinas, Brazil, Jun. 2017.

And more...

Thank you! �Questions?

​

chesteve@�unicamp.br

Limits of human understanding

Technological limits

Hand Crafted Code

Modeling + Code Gen

Slide courtesy of Jonathan Lynam (Ericsson), suggested by Attila Takacs (Ericsson, Inspired by Vinod Khosla @ ONS2014 �further info: https://www.ericsson.com/research-blog/sdn/model-driven-networking-looom/

Hand Crafted:

* Lower apparent barrier to entry

* Domain Experts + programmers = ???

​

​

Speed, Capacity, Features

Complexity

Tools Driven

* Higher initial complexity

* Machinists, not just Tailors + Apprentices

​

​

Goal: keep curve more flat

17

Overlaying Tradeoffs

Slide courtesy: D. Meyer / J. Doyle

Programmability

Performance

18

Layered architectures make Robustness and Evolvability compatible

Source: adapted from D. Meyer / Doyle

Think OS Kernel + Libraries/Drivers

• Plug-in Architectures
• Control Plane Drivers + APIs
• Model-Driven Everything

SDN / NFV Orchestrator?

19

Universal Laws and Architectures

Source: J. Doyle

20

Universal Laws and Architectures

Performance

Programmability

ASIC

NPU

x86

+

+

-

-

21

P3 trade-offs: Programmability, Performance, Portability

:: packets/sec processed

​

(could be normalized per Energy or Cost)

:: ability to (re-)define the dataplane behaviour

:: ability to easily transfer the dataplane logic to a new platform

22

P3 trade-offs: Programmability, Performance, Portability

Programmability

Portability

ideal

Realm of the Impossible?

Performance

+

+

+

How to push the hard limits?

23

Dataplane trade-offs: Programmability, Performance

Performance

Programmability

ASIC NPU CPU

x86

Cavium TX

Ethernet switch

Assuming same use case.

Normalized for chip technology and transistor count.

How big is the difference?

Source: Pongracz G, Molnar L, Kis ZL, Turanyi Z. Cheap silicon: a myth or reality? Picking the right data plane hardware for software defined networking. HotSDN13

+

+

24

P3 trade-offs: Programmability, Performance, Portability

Programmability

Portability

ASIC

NPU

CPU

Performance

+

How portable is the code defining the logic of an ASIC-based dataplane�or the SW-based dataplane on an NPU or CPU?

25

P3 trade-offs: Increasing Performance

• Overcoming the memory wall
• Structured ASICs
• PISA: New generation of switch ASICs
• High-performance FPGAs
• Fotonics & Packet Optical Integration
• Moore’s law goes manycore
• Advances in high-performance (virtualized) networking stacks
• Fast packet Linux I/O (DPDK, PF_RING, Netamp, VPP…)

​

​

Performance

Programmability

Portability

26

P3 trade-offs: Increasing Portability

Performance

Programmability

Portability

• OpenFlow model not enough
• Dataplane abstractions + Open APIs, e.g. ODP
• High-level DSL (and IR?) e.g., P4
• Multi-Architecture Compiler toolchains
• Model-based everything
• e.g., YANG

​

Source: ONF PIF

27

P3 trade-offs: Increasing Programmability

• SDN + NFV
• New Programming languages at diferent layers
• Automation
• New control loops (feat. Analytics/AI)
• aka. Knowledge-Defined Networking (KDN)
• Open Source
• Open source SW, HW (think Open Compute Project)
• APIs APIs APIs
• Model-driven networking (taming complexity)
• https://www.ericsson.com/research-blog/sdn/model-driven-networking-looom/

​

​

​

Performance

Programmability

Portability

28

References / Further Reading

• John Doyle, Universal laws and architecture
• David Meyer, Macro Trends, Architecture, and the Hidden Nature of Complexity (and what does this have to do with SDN?)
• Russ White, Tradeoffs in Network Complexity
• P4: http://p4.org/ for code and specs (Current release is P416)
• A. Capone & C. Cascone, SDN tutorial, IEEE Netsoft 2015
• Jonathan Lynam, Model-driven networking, 2016
• Diego Kreutz et al., "Software-Defined Networking: A Comprehensive Survey." In Proceedings of the IEEE, Vol. 103, Issue 1, Jan. 2015.
• Pattam Gyanesh Patra, Christian Esteve Rothenberg, Gergely Pongrácz. "MACSAD: High Performance Dataplane Applications on the Move". In IEEE 18th International Conference on High Performance Switching and Routing (HPSR), Campinas, Brazil, Jun. 2017.
• Ahmad Rostami, Katia Obraczka, and Christian Esteve Rothenberg. Tutorial on SDN, NFV and Their Role in 5G. In ACM SIGCOMM Tutorials, Florianopolis, Brazil, Aug. 2016.

29

P3 trade-offs: Programmability, Performance, Portability

Performance

Programmability

Portability

30

One SDN to rule them all

Actually not, different reasonable models and approaches to SDN are being pursued

One SDN controller to rule them all, with a discovery app to find them,�One SDN controller to tell them all, on which switchport to bind them.

In the Data Center, where the packets fly.

Source Poem: http://dovernetworks.com/?p=83

Further reading: http://theborgqueen.wordpress.com/2014/03/31/the-legend-of-sdn-one-controller-to-rule-them-all/

�

Different SDN Models

Canonical/Open SDN

Traditional

Hybrid/Broker

Overlay

Compiler

Different SDN Models to Program / Refactor the Stack

Legacy

​

Data Plane

Mgm.APIs

Distributed L2/L3�Control Plane

Managemt�Software

Mgm.

HAL APIs / Drivers

Orchestrator

SDN

OSS/BSS

Mgm. Apps

Network OS / Bare Metal Switches

33