Internet-Draft Abbreviated Title July 2021
Daniel, et al. Expires 11 January 2022 [Page]
Intended Status:
Standards Track
D.H. Daniel
ZTE Corporation
B.T. Bin
ZTE Corporation
P.L. Peng
China Mobile

Computing Delivery in Routing Network


This document drafts a proposal of Computing Delivery in Routing Network which incorporates both computing and networking metrics into the routing polices and enables the network sensing and scheduling computing services based upon traditional networking services. A mechanism of two-class computing power granularity and two segment forwarding is illustrated for end-to-end networking and computing service in the cloud sites, while major networking and computing actors is defined in terms of functionality. An example work flow is demonstrated, and both control plane and data plane solution consideration is proposed.

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

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 11 January 2022.

Table of Contents

1. Introduction

Computing-related services have been provided in such a way that computing resources either are confined within isolated sites (data centers, MECs etc.) without coordination among multiple sites or they are coordinated and managed within specific and closed service systems, while the industry develops into an era in which the computing resources becomes more and more ubiquitous. Therefore substantial benefits in light of both cost and efficiency resulting from scale of economy, would be brought into multiple industries by intelligently and dynamically connecting the distributed computing resources and rendering the coordinated computing resources as a single virtual resource pool. Although it could be achieved in a routing network-agnostic way as pure application-level solution, additional gains could be reaped with the converged solution of computing and networking routing. Some impressive drafts such as [I-D.liu-dyncast-ps-usecases] and [] analyze the benefits of routing related solution, and give the reference architecture and preliminary test results. End applications could be served not only by fine-grained computing services but also fine-grained networking services rather than the best-effort networking services without routing network involved otherwise. The cost is the burden of maintaining and sensing computing resource status in the networking nodes, and it's bear in mind while formulating this proposal and the issue is addressed to a degree the cost could be acceptable. This draft puts forward some considerations of computing delivery in routing network, which proposes a way to optimize routing by two-class computing power granularity and two segments forwarding.

1.1. Requirements Language

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 [RFC2119].

2. Terminology

3. Computing delivery in routing network reference architecture

Routing network is enabled sensing the computing resource and service in the cloud sites and routing the application flow according to both network and computing metrics by a computing delivery in routing network architecture as illustrated in figure 1. The architecture is a horizontal convergence of cloud and network, while the latter maintains the converged resource status and thus is able to achieve an end to end routing and forwarding policy from a perspective of cloud and network resource. PE1 maintains GCRS with a whole picture of the multiple cloud sites, and executes the routing policy for the network segment between PE1 and PE2 or PE3, namely between ingress and egress, while PE2 maintains LCRS with a focus picture of the cloud site where S1 resides, and establishes a connection towards S1. S1 is an active instance of a specific computing service type (CSI). On top of the role of LCR which maintains LCRS, PE2 and PE3 also fulfill the role GCR which maintains GCRS from neighboring cloud sites. P provides traditional routing and forwarding functionality for computing service flow, and remains unaware of any computing-related status as well as CSI encapsulations.

                                                  +--------+        +--------+
                                          +------>|LCR/GCR |------->|   ICS  |
                                          |       +--------+        +--------+
                +--------+    +--------+  |          PE2                S1
                |  GCR   |--->|  CUR   |--+
                +--------+    +--------+  |          PE3                S2
                   PE1            P       |       +--------+        +--------+
                                          +------>|LCR/GCR |------->|   ICS  |
                                                  +--------+        +--------+
                |<------------ Network domain --------------->|<--Computing->|

Figure 1

3.1. Hierarchical granularity routing scheme

Status updates of computing resource and service in the cloud sites stay in a quite broad range from relatively stable service types and overall resource occupation to extremely dynamic capacity changes as well as busy and idle cycle of service instance. It would be a disaster to build all of the status updates in the network layer which would bring overburdened and volatile routing tables.

It should be reasonable to divide the wide range of computing resource and services into different categories with differentiated characteristics from routing perspective. GCRS and LCRS correspond to cross-site domain and local site domain respectively, and GCRS aggregates the computing resource and service status with low update frequency from multiple cloud sites while LCRS focuses only upon the status with high frequency in the local sites. Under this two-granularity scheme, computing-related routing table of GCRS in the GCR remains in a position roughly as stable as the traditional routing table, and the LCRS in the LCR maintains a near synchronized state table of the highly dynamic updates of computing service instances in the local cloud site. Nonetheless, LCRS focusing upon a single and local cloud site is the normal case while upon multiple sites should be exemption if not impossible.

3.2. Two-segment routing and forwarding

When it comes to end to end service flow routing and forwarding, there is an status information gap between GCRS and LCRS, therefore a two-segment mechanism has to be in place in line with the two-granularity routing scheme demonstrated in 3.1. As is illustrated in figure 2, R1as ingress determines the specific service flow's egress which turns out to be R2 according to policy calculation from GCRS. In particular, the CSI from either in-band or out-band is the only index for R1 to calculate and determine the egress, it's highly possible to make this egress calculation in terms of both networking (bandwidth, latency etc) and computing Service Agreement Level. Nevertheless, the two SLA routing optimization could be decoupled to such a degree that the traditional routing algorithms could remain as they are. The convergence of the SLA policies as well as the methods to make GCR aware of the two SLA is out of scope of this proposal.

                +--------+    +--------+    +--------+    +--------+
                |  GCRS  |--->|        |--->|  LCRS  |--->|   ICS  |
                +--------+    +--------+    +--------+    +--------+
                    R1            R
                |<---------- GCRS segment ---------->|<-   LCRS  ->|

Figure 2

When the service flow arrives at R2 which terminates the GCRS segment routing and determines S1 which is the service instance selected according to LCRS maintained at R2. Again CSI is the only index for LCRS segment routing process.

3.3. CSI routing

CSI encapsulated in the headers and maintained in LCRS and GCRS indicates an abstract service type rather than a geographically explicit destination label, thus the routing scheme based upon CSI is actually a two-part and two-layer process in which CSI only indicates the routing intention of user's requested computing service type where routing does not actually materialize in forwarding plane and the explicit routing destination would be determined by LCRS and GCRS. Therefore the actual routing falls within the traditional routing scheme which remains intact.

3.4. Traffic affinity

CSI holds the only semantics of the service type that could be deployed as multiple instances within specific cloud site or across multiple cloud sites, CSI in the destination field is not explicit enough for all of the service flow packets to be forwarded to a specific destination. Traffic affinity has to be guaranteed at both GCR and LCR. Once the egress is determined at GCR, the binding relationship between the egress and the service flow's unique identification (5-tuple or other specifically designed labels) is maintained and the subsequent flow could be forwarded upon this binding table. Likewise LCR maintains the binding relationship between the service flow identification and the selected service instance.

Traffic affinity could be guaranteed by mechanisms beyond routing layer, but they will not be in the scope of this proposal.

4. Computing delivery in routing network architecture work flow

4.1. Computing resource and service update work flow

The full range of computing resource and service status from a specific cloud site is registered at LCR which maintains LCRS in itself and notifies the part of GCRS to remote GCRs where GCRS would be thus maintained and updated. As is illustrated in figure 3,GCR in R1 from site1 and site 2 is updated by R2 and R3, while LCRS of site 1 in R2 is updated by S1 and LCRS of site 2 in R3 is updated by S2. GCRS in R2 and R3 is updated by each other. Edge routers associating with local cloud site establish a mesh network to update the according GCRS among the whole network domain, the computing resource and services in distributed cloud sites thus are connected and could be utilized as a single pool for the applications rather than the isolated islands.

                                                  +--------+        +--------+
                      +---------------------------|LCR/GCR |<-------|   ICS  |
                      |                           +--------+        +--------+
                +-----V--+    +--------+            A R2 |              S1
                |  GCR   |    |  CUR   |            |    |
                +-----A--+    +--------+            | R3 V              S2
                  R1  |           R               +--------+        +--------+
                      +---------------------------|LCR/GCR |<-------|   ICS  |
                                                  +--------+        +--------+
                |<--------- GCRS update domain ----------->|<-----LCRS------>|

Figure 3

4.2. Service flow routing and forwarding work flow

From perspective of the service work flow, more details have actually been demonstrated in 3.2 and 3.3. Rather than the traditional destination-oriented routing mechanism and the segment routing in which the ingress router is explicitly aware of a specific destination, CSI as an abstract label without semantics of physical address works as the required destination from viewpoint of the user in computing delivery in routing network architecture. Therefore the service flow has to be routed and forwarded segment by segment in which the two segment destinations are determined by GCRS and LCRS respectively.

5. Control plane

5.1. Centralized control plane

LCRS's volatility makes it infeasible to be maintained and controlled in a centralized entity, GCRS is the chief computing resource and service status information to be collected and managed in the controller when it comes to centralized control plane with regard to computing delivery in routing network architecture. Routing and forwarding policies from GCRS calculated in the centralized controller, as is demonstrated in 3.2, apply only to the segment from ingress and egress, while the second segment routing policy from egress to the selected service instance in the cloud site is determined by LCRS at egress.

Hierarchically centralized control plane architecture would be strongly recommended under the circumstances of nationwide network and cloud management.

5.2. Distributed control plane

GCRS is updated among the edge routers which have been connected in a mesh way that each pair of edge routers could exchange GCRS to each other, while LCRS will be unidirectionally updated from cloud site to the associated edge router in which LCRS is maintained and its update process is terminated.

Protocol consideration upon which GCRS and LCRS is updated is out of the scope of this proposal.

5.3. Hybrid control plane

It should be more efficient to update the GCRS by a distributed way than a centralized way in terms of routing request and response in a limited network and cloud domain, but be the opposite case in a nationwide circumstance. This is how hybrid control plane could be deployed in such a scheme that overall optimization is achieved.

6. Data plane

6.1. CSI encapsulation

Computing service identification is the predominant index across the entire computing delivery in routing network architecture under which a new virtual routing scheme is employed with CSI working as the virtual destination. Data plane indicates the routing and forwarding orientation with CSI by inquiring GCRS and LCRS at GCR and LCR respectively. CSI encapsulation could be achieved by extending the existing packet header and also achieved by designing a dedicated shim layer, which along with the specific structure of CSI are out of the scope of this proposal.

6.2. CSI for GCR, CUR and LCR

GCR encapsulates CSI in a designated header format as a proxy by translating the user-originated CSI format, and makes the first segment routing policy and starts routing and forwarding the service traffic. CUR ignores CSI and simply forwards the traffic as usual. LCR decapsulates CSI and makes the second segment routing policy and completes the last hop routing and forwarding.

7. Summary

It would significantly benefit the industry by connecting and coordinating the distributed computing resources and services and more so by further converging networking and computing resource. Uncertainty and the potential impacts over the ongoing network architecture is the main reason for the community to think twice. By classifying the end to end routing and forwarding path into two segments, the impacts from computing status and metrics are to be reduced to a degree they would be as acceptable and comfortable enough as they are as networking status and metrics. In particular, employment of CSI in computing delivery in routing network architecture enables a new service routing possibility perfectly compatible with the ongoing routing architecture.

8. Acknowledgements

To be added upon contributions, comments and suggestions.

9. IANA Considerations

This memo includes no request to IANA.

10. Security Considerations

As information originated from the third party (cloud sites), both GCRS and LCRS would be frequently updated in the network domain, both security threats against the routing mechanisms and credibility and security issues of the computing services should be taken into account by architecture designing. The detailed analysis as well as solution consideration will be proposed in the updated version of the draft.

11. Informative References

Li, Y., "Dynamic-Anycast Architecture", , <>.
Liu, Peng., "Dynamic-Anycast (Dyncast) Use Cases and Problem Statement", , <>.
Bradner, S., "Key words for use in RFCs to Indicate Requirement Levels", BCP 14, RFC 2119, DOI 10.17487/RFC2119, , <>.

Authors' Addresses

Daniel Huang
ZTE Corporation
Bin Tan
ZTE Corporation
Peng Liu
China Mobile