Skip to content

Campus Main — IS Calculations at Every Layer (B5 to B9)

Campus site profile

Parameter Value Notes
Human agents 200 Split across 2 access switches (100 per switch)
IoT / edge sensors 500 Cameras, sensors, smart building
GPU inference pods 10 In-house DC — Tier 3 + Tier 4
U_eff (200×1.0) + (500×0.2) + (10×7.0) = 370 200 + 100 + 70 = 370
AIW 12.0 Mbps per load unit Full multimodal: STT + LLM + screen + RAG + fraud + burst
CS 4.5 DPDP + PCI DSS + high-risk traffic + proprietary AI models
LL 4.5 Latency-sensitive voice AI + fraud detection (~25ms budget)
Availability 99.9% required Triple-path WAN + in-house DC
IS numerator 370 × 12.0 × 4.5 × 4.5 = 89,910 Fixed — used in all campus calculations

CS = 4.5 and LL = 4.5 — architecture implications

CS = 4.5 means data cannot leave the enterprise network without full DLP inspection. LL = 4.5 means the RTT budget is approximately 25ms — impossible over any cloud WAN path from India. Both values mandate in-house DC inference for real-time AI workloads. These are not preferences — they are hard constraints.

What is B here?

The 25G uplink (or port-channel) from each campus access switch to the campus core distribution layer. Each switch serves 100 agents plus 250 IoT devices. The 99.9% availability requirement mandates dual uplinks.

Per-switch U_eff calculation

Scope: 100 agents + 250 IoT per switch (half of campus total)
U_eff_per_switch = (100 × 1.0) + (250 × 0.2) = 100 + 50 = 150
Per-switch numerator = 150 × 12.0 × 4.5 × 4.5 = 36,450

B   = 25,000 Mbps  (25G uplink)
A   = 0.97         (dedicated campus fabric uplink)

IS  = 36,450 / (25,000 × 0.97)
IS  = 36,450 / 24,250
IS  = 1.50   ← MONITOR

Failover (this is also the failover IS — single link must carry full load):
IS  = 1.50 — acceptable but tight

Effective B = 50,000 Mbps  (dual 25G LACP port-channel)
A           = 0.97

IS  = 36,450 / (50,000 × 0.97)
IS  = 36,450 / 48,500
IS  = 0.75   ← OPTIMAL

Failover (one 25G link fails, other carries full 150-unit load):
IS  = 36,450 / (25,000 × 0.97) = 1.50  ← Monitor — service maintained

Recommendation: Dual 25G port-channel. IS = 0.75 normal, IS = 1.50 failover. Mandatory for 99.9% availability — dual uplinks ensure a single link failure does not push IS above 3.

10G uplink comparison:
IS = 36,450 / (10,000 × 0.97) = 3.75  ← UPGRADE REQUIRED

10G access uplinks are insufficient for a 100-agent campus floor with
full multimodal AI at AIW = 12 Mbps. Minimum 25G required.

Bandwidth consumed at B8

Actual AI traffic per switch = 150 × 12.0 = 1,800 Mbps = 1.8 Gbps
Plus east-west traffic (GPU pod → agents): approximately 600 Mbps additional
Total: ~2.4 Gbps per switch uplink

Utilisation on 25G: 2.4 / 25 = 9.6% (single link)
Utilisation on 50G PC: 2.4 / 50 = 4.8% (normal operation)

B9 — Campus Core to In-House DC (100G internal fabric)

What is B here?

The 100G link from the campus core switch to the in-house data centre housing the GPU inference pods. All AI inference traffic — agent assist, STT, fraud detection — flows through this link from agents to GPUs and back. This is the internal AI fabric path.

Calculation

Scope: Full campus (all 370 load units use the DC)
U_eff = 370, Numerator = 89,910
B   = 100,000 Mbps  (100G dedicated internal link)
A   = 0.99           (dedicated dark fibre, local DC, no contention)

IS  = 89,910 / (100,000 × 0.99)
IS  = 89,910 / 99,000
IS  = 0.91   ← OPTIMAL

Actual bandwidth:    370 × 12.0 = 4,440 Mbps = 4.4 Gbps
Utilisation:         4.4 / 100 = 4.4%  (95.6% spare capacity)
At full burst (×2.5): 11 Gbps / 100G = 11% — still well within optimal

IS = 0.91 — Optimal. The core-to-DC link is never the AI bottleneck. 95.6% spare capacity.

Why dual 100G?

Not for IS — single 100G is more than sufficient. Dual 100G is for availability:

Dual 100G port-channel:
IS = 89,910 / (200,000 × 0.99) = 0.45  ← over-provisioned

Add a second 100G only if the 99.9% availability requirement demands that the DC link survives a fibre cut. For most campus designs, dual 25G port-channels (50G aggregate) to the DC are sufficient.

Don't over-invest in B9

The internal DC fabric is never the bottleneck. Upgrade investment should go to WAN circuits and edge AI infrastructure, not to the internal 100G fabric. A single 100G link to the in-house DC is architecturally sound even at 2× the campus load.

B5 — MPLS 10G WAN (Primary Campus WAN)

What is B here?

The 10G MPLS Committed Information Rate (CIR) from campus to cloud regions and other enterprise sites. MPLS provides a guaranteed SLA and managed QoS — A = 0.92.

The key variable is not the MPLS circuit size but how much AI traffic crosses it. This depends entirely on MCP tiering.

AIW sent to cloud (Mbps) CS on cloud LL on cloud IS on MPLS 10G Status
12.0 — all cloud, no edge AI 4.5 4.5 9.77 Blocker
6.0 — 50% workloads edge 3.5 3.5 4.26 Upgrade needed
3.0 — 75% workloads edge 2.5 2.5 1.52 Monitor
2.0 — MCP Tier ½ only 2.0 2.0 0.32 Optimal
1.0 — heavy edge AI 1.5 1.5 0.09 Optimal

Post-MCP calculation (production design)

Cloud-only workloads (Tier 1/2): report gen + screen analytics
AIW_cloud = 2.0 Mbps per agent
CS_cloud  = 2.0 (non-PII, non-regulated cloud queries)
LL_cloud  = 2.0 (batch tolerant, 200ms acceptable)

IS_MPLS = (370 × 2.0 × 2.0 × 2.0) / (10,000 × 0.92)
IS_MPLS = 2,960 / 9,200
IS_MPLS = 0.32   ← OPTIMAL

WAN traffic reduction:  12.0 → 2.0 Mbps/agent = 83% less WAN traffic
Cloud egress reduction: 100% → 18% of total AI flows

MPLS QoS configuration

Because MPLS supports end-to-end QoS with provider enforcement, configure these DSCP markings for cloud-bound traffic:

Class           DSCP    Queue    Max share    AI function
─────────────────────────────────────────────────────────
RAG queries     CS4     High     20%          Knowledge base
Analytics       CS3     Medium   35%          Sentiment, batch
Report gen      CS2     Medium   25%          Document AI
Model sync      CS1     Low      10%          Off-peak only
Background      BE      Default   5%          Unclassified

MPLS QoS passthrough

Verify with your MPLS provider that DSCP markings are honoured end-to-end. Some providers re-mark traffic at the provider edge. Request the provider's DSCP-to-class mapping and confirm AI traffic (CS4 and above) receives the promised latency treatment.

B6 — Internet 5G ISP (Secondary Campus WAN)

What is B here?

A 5G ISP link (fibre-connected 5G infrastructure, not cellular) providing up to 5 Gbps. Used as the active-active secondary path alongside MPLS in the SD-WAN design. A = 0.72 — shared internet medium, no end-to-end SLA.

Normal operation (active-active with MPLS)

Cloud traffic only (post-MCP):
U_eff=370, AIW=2.0, CS=2.0, LL=2.0

Internet 5G ISP alone (B=5,000, A=0.72):
IS = (370 × 2.0 × 2.0 × 2.0) / (5,000 × 0.72)
IS = 2,960 / 3,600
IS = 0.82   ← OPTIMAL

SD-WAN aggregate (MPLS 10G + Internet 5G, A=0.85):
IS = 2,960 / (15,000 × 0.85)
IS = 2,960 / 12,750
IS = 0.23   ← OPTIMAL

Failover scenario: MPLS fails

Internet 5G ISP carries entire cloud AI load alone:
IS = 0.82   ← OPTIMAL

Service is fully maintained. IS does not exceed threshold even in failover.
This confirms the 99.9% availability design is valid for cloud-bound AI.

What the internet path cannot do

The internet (B6) cannot carry LL = 4 or LL = 5 workloads even at IS = 0.82, because the RTT constraint is violated regardless of bandwidth:

RTT to nearest cloud (Mumbai AWS): 8–15ms
+ Firewall/LB: 5ms
+ Inference in cloud: 25–40ms
+ Return path: 8–15ms
Total RTT: 46–75ms

LL = 4 budget: 31ms
LL = 5 budget: 20ms

Cloud AI via internet fails LL = 4 and LL = 5 on physics, not bandwidth.
These workloads must run in the in-house DC regardless of WAN bandwidth.

B7 — 5G Cellular Backup (Tertiary Emergency Path)

What is B here?

A 5G cellular modem providing realistic sustained throughput of 300–600 Mbps (theoretical maximum 1+ Gbps, derated for enterprise planning). A = 0.60 — wireless medium, shared tower infrastructure, no enterprise SLA.

Emergency failover calculation

Assumptions for emergency mode:
- Both MPLS and internet fail simultaneously (extremely rare)
- 50% of agents remain active: U_eff = 185
- Only cloud analytics traffic (Tier 1/2): AIW = 2.0, CS = 2.0, LL = 2.0

5G at 500 Mbps sustained:
IS = (185 × 2.0 × 2.0 × 2.0) / (500 × 0.60)
IS = 1,480 / 300
IS = 4.93   ← UPGRADE ZONE — acceptable for emergency window

5G at 1 Gbps (good signal):
IS = 1,480 / 600
IS = 2.47   ← MONITOR — acceptable emergency mode

Critical AI (fraud detect, agent assist):
→ Running from in-house DC — ZERO WAN dependency
→ These continue at full performance regardless of 5G status

Availability mathematics

MPLS availability:     99.5%  → outage probability: 0.005
Internet availability: 99.2%  → outage probability: 0.008
5G cellular:           99.8%  → outage probability: 0.002

Triple-path combined outage = 0.005 × 0.008 × 0.002 = 0.00000008
Availability = 1 - 0.00000008 = 99.9999%

Requirement: 99.9%
Achieved:    99.9999%  ← exceeds by 1,000×

The triple-path design (MPLS + Internet + 5G) with in-house DC edge AI far exceeds the 99.9% requirement. The 5G cellular modem is the insurance policy, not the operational path.

5G cellular configuration for SD-WAN

SD-WAN policy for 5G:
  Priority:           3 (tertiary — only when both other paths fail)
  Trigger:            Both MPLS and internet fail simultaneously
  Traffic allowed:    CS1–CS3 only (analytics, reports, non-time-critical)
  Traffic blocked:    EF, CS5, CS4 (these run from in-house DC anyway)
  Scheduler:          No model sync on 5G (too slow, too expensive)
  Max link usage:     80% of 5G capacity

Summary: campus IS across all layers

Layer B value A IS (all-cloud naive) IS (optimised) Bottleneck?
B8 — Access uplink (single 25G) 25G 0.97 1.50 1.50 No
B8 — Access uplink (dual 25G PC) 50G 0.97 0.75 0.75 No
B9 — Core to DC 100G 100G 0.99 0.91 0.91 No
B5 — MPLS 10G 10G 0.92 9.77 0.32 Yes (solved by MCP)
B6 — Internet 5G ISP 5G 0.72 24.9 0.82 Yes (solved by MCP)
B7 — 5G cellular 500M 0.60 — 2.47 Emergency only

The LAN and internal DC fabric are always fine. The WAN fails without MCP tiering. MCP tiering — not bandwidth upgrades — is the solution.