Network Slicing with SRv6¶
Network slicing creates isolated logical networks over a shared physical infrastructure, each with its own topology, SIDs, QoS guarantees, and service-level objectives. SRv6 enables slicing natively through the combination of Flex-Algorithm, QoS policies, and SR Policy color steering — no additional overlay protocols required.
What Is Network Slicing?¶
A network slice is an end-to-end logical partition that provides:
- Topology isolation — each slice uses a different set of paths through the network
- Resource guarantees — dedicated bandwidth, latency, or jitter targets per slice
- Independent policies — each slice can have its own TE, protection, and routing constraints
Slicing vs VPN vs Traffic Engineering¶
| Concept | Layer | Isolation Type | Mechanism |
|---|---|---|---|
| VPN (VRF) | L3 | Routing table isolation | Separate FIBs per customer |
| Traffic Engineering | Path | Path selection | SR Policy / RSVP-TE |
| Network Slice | Topology + Resources | Topology + bandwidth + QoS | Flex-Algo + QoS + Color |
VPNs ride on slices
Network slicing and VPNs are complementary. A customer VPN can be mapped onto a specific slice (e.g., "Gold slice") to get both routing isolation and transport-level SLA guarantees.
SRv6 Slicing Building Blocks¶
SRv6 network slicing combines three mechanisms:
graph TB
subgraph Slice Architecture
FA[Flex-Algorithm<br/>Topology Slicing] --> SLICE[Network Slice]
QOS[QoS Policy<br/>Resource Slicing] --> SLICE
COLOR[SR Policy + Color<br/>Service Steering] --> SLICE
end
SLICE --> S1[Slice 1: Low Latency]
SLICE --> S2[Slice 2: High Bandwidth]
SLICE --> S3[Slice 3: Best Effort] 1. Flex-Algorithm (Topology Slicing)¶
Each slice maps to a Flex-Algorithm (128-255), which computes an independent topology based on constraints like latency, affinity, or TE metric. Every node in the slice advertises a per-algorithm locator and SID.
| Slice | Flex-Algo | Constraint | Effect |
|---|---|---|---|
| Low Latency | 128 | Minimize delay metric | Uses only low-latency links |
| High Bandwidth | 129 | Minimize TE metric + affinity | Avoids congested links |
| Best Effort | 0 (default) | Minimize IGP metric | Uses all links |
For full details, see Flex-Algorithm.
2. QoS (Resource Slicing)¶
Flex-Algorithm provides topology isolation, but bandwidth guarantees require QoS policies:
- DSCP marking per slice — each slice gets a distinct DSCP value
- Per-slice scheduling — strict priority, weighted fair queuing, or bandwidth reservation
- Policers — rate limiting per slice to enforce SLA boundaries
For full details, see QoS.
3. SR Policy + Color (Service Steering)¶
BGP color extended communities map services into the correct slice:
- Service route is advertised with a color community (e.g., color 128 = low-latency slice)
- Headend PE matches the color to an SR Policy or Flex-Algo SID
- Traffic is steered into the slice automatically
For full details, see SR Policy.
Use Case 1: 5G Network Slicing¶
3GPP defines standardized slice types that map directly to SRv6 Flex-Algorithm slices:
| 3GPP Slice Type | SST | Description | SRv6 Mapping |
|---|---|---|---|
| eMBB | 1 | Enhanced Mobile Broadband | Flex-Algo 129 (bandwidth-optimized) |
| URLLC | 2 | Ultra-Reliable Low Latency | Flex-Algo 128 (latency-optimized) |
| mMTC | 3 | Massive Machine-Type Comm. | Flex-Algo 0 (best effort) |
End-to-End 5G Slice¶
graph LR
RAN[RAN / gNB] -->|URLLC: Algo 128| TR1[Transport PE]
RAN -->|eMBB: Algo 129| TR1
TR1 -->|Slice-specific path| CORE((SRv6 Core))
CORE --> UPF[UPF / Core PE] The SRv6 MUP (Mobile User Plane) behaviors from RFC 9433 integrate with Flex-Algo slicing to provide end-to-end slice awareness from RAN to core.
For the full 5G transport architecture, see 5G Transport.
Use Case 2: Enterprise Network Slicing¶
Enterprise operators use slicing to offer tiered services with guaranteed SLAs:
| Tier | Flex-Algo | DSCP | QoS Policy | SLA Target |
|---|---|---|---|---|
| Gold | 128 | EF (46) | Strict priority, 30% BW guarantee | <10ms latency, 99.999% |
| Silver | 129 | AF31 (26) | Weighted fair queue, 50% BW | <50ms latency, 99.99% |
| Bronze | 0 | BE (0) | Best effort, remaining BW | Best effort |
Configuration Example¶
!! Flex-Algorithm definitions
router isis CORE
flex-algo 128
metric-type delay
advertise-definition
!
flex-algo 129
affinity exclude-any CONGESTED
advertise-definition
!
!
!! Per-slice locators
segment-routing
srv6
locators
locator GOLD
micro-segment behavior unode psp-usd
prefix fc00:128:1::/48
algorithm 128
!
locator SILVER
micro-segment behavior unode psp-usd
prefix fc00:129:1::/48
algorithm 129
!
locator BEST-EFFORT
micro-segment behavior unode psp-usd
prefix fc00:0:1::/48
!
!
!
!
!! Color steering for VPN → slice mapping
route-policy SET-GOLD-COLOR
set extcommunity color 128
end-policy
router bgp 65000
vrf GOLD-CUSTOMER
address-family ipv4 unicast
route-policy SET-GOLD-COLOR out
!
!
!
# Flex-Algo definition
set protocols isis flex-algorithm 128 metric-type min-delay
set protocols isis flex-algorithm 129 affinity exclude CONGESTED
# Per-slice locators
set routing-options segment-routing-v6 locator GOLD fc00:128:1::/48 algorithm 128
set routing-options segment-routing-v6 locator SILVER fc00:129:1::/48 algorithm 129
set routing-options segment-routing-v6 locator BEST-EFFORT fc00:0:1::/48
# Color steering
set policy-options community GOLD-COLOR color 128
set policy-options policy-statement GOLD-EXPORT term 1 then community add GOLD-COLOR
Use Case 3: Data Center Slicing¶
In CLOS fabrics, Flex-Algorithm slicing isolates different traffic types:
| Slice | Purpose | Constraint |
|---|---|---|
| GPU Fabric | AI/ML training traffic | Low-latency links, RDMA-capable paths |
| Storage | NVMe-oF, iSCSI | High-bandwidth, redundant paths |
| Management | SSH, SNMP, telemetry | Best-effort, any path |
Each slice gets its own Flex-Algo and locator, ensuring that GPU traffic never contends with storage replication on the same path.
For CLOS fabric design, see CLOS & Load Balancing. For AI/ML-specific networking, see AI/ML Networks.
Slice Design Considerations¶
SID Allocation per Slice¶
Each slice requires a dedicated locator block. Plan allocations carefully:
| Slice | Algorithm | Locator Block | Nodes |
|---|---|---|---|
| Best Effort | 0 | fc00:0::/32 | All |
| Low Latency | 128 | fc00:128::/32 | All |
| High Bandwidth | 129 | fc00:129::/32 | All |
| Custom | 130 | fc00:130::/32 | Subset |
Monitoring and SLA Assurance¶
Per-slice monitoring is essential for SLA compliance:
- SRPM (SR Performance Measurement) — per-slice delay/loss probes using slice-specific SIDs. See SRPM.
- Telemetry — per-slice counters on interfaces and forwarding entries. See Telemetry.
- IPFIX — per-slice flow records using RFC 9487 SRv6 Information Elements.
Scalability¶
| Factor | Limit | Notes |
|---|---|---|
| Flex-Algorithms | 128 (128-255) | Practical limit: 4-8 slices per network |
| Locators per node | Platform-dependent | Typically 8-16 concurrent locators |
| FIB impact | N x nodes | Each algorithm multiplies the IPv6 FIB by the number of participating nodes |
FIB scaling
Each additional Flex-Algorithm adds a full set of locator routes to the FIB. With 1000 nodes and 4 algorithms, expect ~4000 locator entries. Plan hardware FIB capacity accordingly.
Comparison: SRv6 Slicing vs Alternatives¶
| Aspect | VRF-Only | MPLS TE Tunnels | FlexE | SRv6 Flex-Algo |
|---|---|---|---|---|
| Isolation | Routing only | Path only | Physical (TDM) | Topology + path |
| Complexity | Low | High (RSVP state) | Medium (hardware) | Low-medium |
| Scalability | High | Limited (~10K tunnels) | Low (port-based) | High |
| SLA granularity | Coarse (per VRF) | Fine (per tunnel) | Exact (hard isolation) | Fine (per algorithm) |
| Automation | Limited | Manual | Manual | Color steering + ODN |
| Hardware | Any | Any | FlexE-capable | Any SRv6-capable |
Further Reading¶
- Flex-Algorithm — Constraint-based topology computation
- QoS — Per-slice quality of service policies
- SR Policy — Color steering and ODN for slice assignment
- 5G Transport — 3GPP slicing with SRv6 transport
- AI/ML Networks — GPU fabric slicing
- SD-WAN — SD-WAN overlay with network slicing
References¶
- RFC 9350 - IGP Flexible Algorithm - Defines Flex-Algo for constraint-based path computation, the foundation of SRv6 network slicing
- RFC 9256 - SR Policy Architecture - SR Policy framework including color steering for mapping services to slices
- 3GPP TS 23.501 - System architecture for the 5G System - 5G system architecture defining network slice types (eMBB, URLLC, mMTC)
- RFC 9433 - SRv6 for the Mobile User Plane - MUP behaviors integrating with network slicing for 5G
- draft-ietf-teas-ietf-network-slices - IETF framework for network slicing