Quality of Service for ICN in the IoT
draft-gundogan-icnrg-iotqos-01

Abstract

This document describes manageable resources in ICN IoT deployments and a lightweight traffic classification method for mapping priorities to resources. Management methods are further derived for controlling latency and reliability of traffic flows in constrained environments. This work includes a distributed management of the heterogeneous resources (i) forwarding capacities, (ii) Pending Interest Table (PIT) space, and (iii) in-network data storage. By correlating these common ICN resources, performance measures can be optimized without sacrificing concurrent traffic demands. Different from the IP world, QoS in ICN can be benifical for all participants and manage 'quality instead of unfairness'.

Status of This Memo

This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.

Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.

Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."

This Internet-Draft will expire on January 9, 2020.

Copyright Notice

Copyright (c) 2019 IETF Trust and the persons identified as the document authors. All rights reserved.

This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document.

Table of Contents

• 1. Introduction
• 2. Terminology
• 3. Manageable Resources in the IoT
• 3.1. Link Layer
• 3.2. Pending Interest Table
• 3.3. Content Store
• 4. Traffic Flow Classification
• 5. Priority Handling
• 6. Distributed QoS Management
• 6.1. Locally Isolated Decisions
• 6.2. Local Resource Correlations
• 6.3. Distributed Resource Coordination
• 7. Implementation Report and Guidance
• 8. Security Considerations
• 9. IANA Considerations
• 10. References
• 10.1. Normative References
• 10.2. Informative References
• Acknowledgments
• Authors' Addresses

1. Introduction

The performance of networked systems is largely determined by the resources available for forwarding messages between components. In addition to link capacities and buffer queues, Information-centric Networks rely on additional resources that shape its overall performance, namely Pending Interest Table space, and caching capacity.

Typical IoT deployments add tight resource constraints to this picture [RFC7228]: Nodes have processing and memory limitations, the underlying link layer technologies are lossy and restricted in bandwidth. Particularly in multi-hop networks, such constraints affect the overall performance, create bottlenecks, but may lead to cascading packet loss or energy depletion when PIT resources are independently evicted and forwarding states decorrelate [DECORRELATION]. Overprovisioning to counter performance flaws is infeasible for many IoT scenarios as it inflicts with use cases and increases deployment costs. Quality of Service (QoS) is a method to enhance overall performance by redistributing resources to a subset of messages, and - in the constrained IoT use case - to coordinate operations under resource scarcity.

IoT applications follow various use cases, of which different QoS requirements can be derived. While periodic sensor readings often comply with unmanaged performance, industrial control messaging or security alerts require (very) low latency, and safety-critical environmental recording or network management need (highly) reliable network services. Both quality levels can only be assured by appropriate resource reservations.

In order to achieve a QoS-aware information-centric IoT deployment, this document describes manageable resources in Section 3, defines a flow classification method in Section 4, and specifies priorities and their mappings in Section 5.

2. Terminology

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in RFC 2119 [RFC2119]. The use of the term, "silently ignore" is not defined in RFC 2119. However, the term is used in this document and can be similarly construed.

This document uses the terminology of [RFC7476], [RFC7927], and [RFC7945] for ICN entities.

The following terms are used in the document and defined as follows:

Traffic Flow

A traffic flow is a sequence of messages (Interest and data) that belong to one specific communication context. Due to in-network caching, ICN flows may be delocalized. A flow may also relate to several requesters in the presence of Interest aggregation.

3. Manageable Resources in the IoT

The following resources contribute to the overall network performance in Information-Centric IoT Networking and need management for QoS control.

3.1. Link Layer

The link layer manages access to the media and provides space to buffer packets. Low latency applications require that requests are prioritized compared to regular priority data. Based on the request response pattern of ICN, link layer resources can be preallocated for data packets.

3.2. Pending Interest Table

The Pending Interest Table (PIT) stores open requests at each hop. PIT resources are allocated when requests are forwarded, and they are released on returning responses.

Placement and replacement strategies of PIT entries directly influence the latency and reliability properties of traffic flows and thus should consider prioritization schemes. If the PIT is not saturated new PIT entries can be added. If the PIT is saturated, requests with higher priority should replace requests with lower priority to prevent higher latencies due to retransmissions.

3.3. Content Store

Content stores (CS) enable transparent in-network caching and thus improve the transport in wireless and lossy environments by reducing hop traversals for content requests [NDN-EXP].

Placement and replacement strategies of data in content stores can affect the latency and reliability properties of traffic flows. The latency can be reduced by placing data closer to the consumers. Reliability can be improved by replicating data in multiple content stores to be resilient to node failures.

4. Traffic Flow Classification

This document defines a traffic flow classification mechanism that aggregates names into equivalence classes in order to apply resource allocation decisions on messages of particular traffic flows.

A traffic class is a name prefix and each device maintains a list of valid classes. The actual distribution of traffic classes is out of scope of this document. The classes for request and response messages are derived by performing a longest prefix match (LPM) with the list of valid traffic classes and the Name TLV of the message. Examples are given in Figure 1.

list =
["/org", "/org /Hamburg", "/org /Berlin", "/org /Berlin /sensor" ]

LPM("/com"                      ,list) = ""
LPM("/org /Germany"             ,list) = "/org"
LPM("/org /Hamburg"             ,list) = "/org /Hamburg"
LPM("/org /Berlin /sensor /temp",list) = "/org /Berlin /sensor"

Figure 1: Example traffic flow class matches.

The empty traffic class "" is the default class for messages that do not match any valid traffic classes in the class list.

5. Priority Handling

We define two priority levels to set the priorities for traffic flows in regards to latency and reliability.

• Latency:
  • PROMPT
  • REGULAR
• Reliability:
  • RELIABLE
  • REGULAR

Each list entry of the traffic class list from Section 4 has an associated priority tuple which distinctly specifies priority levels for the resources in Section 3. The tuple is of the following form:

priority tuple = < LATENCY_PRIORITY, RELIABILITY_PRIORITY >

Figure 2: Schema of the priority tuple.

6. Distributed QoS Management

The mechanisms used to achieve QoS management is divided into three classes, depending on the level of interdependency exhibited between mechanisms on the same device or between devices.

6.1. Locally Isolated Decisions

This class includes decisions that have no interaction with other mechanisms on the local or other devices.

Prioritized Forwarding:

As described above, the link layer provides space to buffer outgoing packets. For the two latency priorities, this space can be allocated into the following two queues:

• PROMPT_FORWARDING_QUEUE
• REGULAR_FORWARDING_QUEUE

Packets will be appended to the queue corresponding to their priority level.

Caching Decisions:

The decisions to cache content obey the priority order reliable to regular. In the presence of probabilistic caching strategies, the weights are set accordingly.

PIT Management:

For saturated PITs, the management operates in favor of rapid packet forwarding, so prompt Interests replace regular Interests. Newly arriving Interests that meet a PIT with saturated entries of equal or higher priority are dropped.

6.2. Local Resource Correlations

These are decisions that entail interaction between mechanisms on the same device (intra-device correlations). This includes the correlation between the caching decision and cache replacement strategies.

• If arriving Data meets a valid PIT entry, Data is forwarded according to priorities. Reliable Data is cached with priority. In the case of exhausted prioritized forwarding queues, prompt traffic is cached with the highest priority, because Interest retransmissions are likely to occur.
• If arriving Data meets no valid PIT entry, caching follows the order prompt (highest) to regular (lowest). For probabilistic caching, weights are adjusted correspondingly.

6.3. Distributed Resource Coordination

These decisions affect resources across multiple or all devices in the network (inter-device correlations). These include maintaining PIT coherence by ensuring that all nodes apply uniform QoS mechanisms when replacing content of different service classes, as well as achieving CS diversity by introducing probabilistic caching based on priority classes. In this document, distributed coordination is achieved as follows:

PIT Coherence:

Coherence is increased by applying the same PIT eviction strategy at all nodes. In this case, evict regular before reliable before prompt.

Cache Efficiency:

Efficiency increases with probabilistic caching using the coordination of equal cache weights. The use of probabilistic caching reduces the risk of starvation for low priority content, even if high priority flows dominate the network.

7. Implementation Report and Guidance

The proposed resource management methods have been implemented as an extension of the NDN/CCNx software stack [CCN-LITE] in its IoT version on RIOT [RIOT].

Constrained memory and cpu resources limit the use of an elaborate prioritized buffer queue management. With these constraints, IoT nodes usually employ forwarding queues that can only hold one to two packets at once. Despite these challenges, the proposed methods show visible improvements on forwarding delays.

Experimental evaluations will be added in this section that show the implications of unmanaged PIT and CS resources for traffic forwarding in a resource-constrained environment.

8. Security Considerations

TODO

9. IANA Considerations

TODO

10. References

10.1. Normative References

10.2. Informative References

Acknowledgments

This work was stimulated by fruitful discussions in the ICNRG research group. We would like to thank all active members for constructive thoughts and feedback. In particular, the authors would like to thank Ilya Moiseenko and Dave Oran for their work provided in [I-D.moiseenko-icnrg-flowclass]. This work was supported in part by the German Federal Ministry of Research and Education within the I3 project.

Authors' Addresses