Normalized Delivery Time (NDT)

Why a New Metric?

Delivery time depends on SNR: at higher SNR you deliver faster. To compare cache-aided and baseline architectures fairly, we need to normalize out the SNR dependence. The result is the normalized delivery time (NDT), Ξ”\Delta, which measures delivery time in units of the reference baseline time. NDT = 1 means the cache- aided architecture delivers as fast as an infinite-fronthaul baseline; NDT > 1 means slower; NDT < 1 is impossible (baseline is the best possible).

This metric, introduced by Sengupta-Tandon-Simeone (2017) and refined by the CommIT group in many settings, is the workhorse of C-RAN caching analysis. It isolates the cache-fronthaul tradeoff from the underlying SNR and channel details.

Definition:

Normalized Delivery Time

For a cache-aided C-RAN with delivery time T(SNR)T(\text{SNR}) and reference delivery time Tref(SNR)=K/log⁑2SNRT_\text{ref}(\text{SNR}) = K/\log_2 \text{SNR} (baseline single-user MU-MIMO), the normalized delivery time is Ξ”(M,CF)β€…β€Šβ‰œβ€…β€Šlim⁑SNRβ†’βˆžT(SNR)Tref(SNR).\Delta(M, C_F) \;\triangleq\; \lim_{\text{SNR} \to \infty} \frac{T(\text{SNR})}{T_\text{ref}(\text{SNR})}. NDT = 1 means "delivers as fast as MU-MIMO with infinite fronthaul." NDT > 1 means "slower β€” the fronthaul/cache bottleneck dominates." NDT < 1 is impossible.

The NDT is a function of two key parameters: the per-EN cache MM and the per-EN fronthaul CFC_F.

NDT is a dimensionless ratio, analogous to the DoF in Chapter 5 β€” it isolates the scaling-invariant tradeoff. At high SNR, the absolute delivery time is Ξ”β‹…Tref\Delta \cdot T_\text{ref}, capturing both the SNR-dependent link rate and the architectural overhead.

Theorem: Fundamental NDT Bounds

For a cloud-RAN with NENN_\text{EN} ENs, cache MM per EN, fronthaul CFC_F per EN, KK users, and library NN: max⁑ ⁣(1,β€…β€ŠK(1βˆ’ΞΌ)NENCF+K(1βˆ’ΞΌ)NENβ‹…min⁑(LEN,K))≀Δ(M,CF)≀?\max\!\left(1,\; \frac{K(1-\mu)}{N_\text{EN} C_F} + \frac{K(1-\mu)}{N_\text{EN} \cdot \min(L_\text{EN}, K)}\right) \leq \Delta(M, C_F) \leq ? Roughly: NDT=max⁑(1,K(1βˆ’ΞΌ)/(NENC)+K(1βˆ’ΞΌ)/(NEN))\mathrm{NDT} = \max(1, K(1-\mu)/(N_{\text{EN}} C) + K(1-\mu)/(N_{\text{EN}})), separating the fronthaul-limited term from the downlink-limited term.

Delivery time is governed by two bottlenecks: (i) the fronthaul must carry (1βˆ’ΞΌ)K(1-\mu) K files' worth of data to the ENs; (ii) the downlink must carry the same amount to users. NDT is dominated by the slower.

When Cβ†’βˆžC \to \infty: NDT = K/NEN/LENK/N_\text{EN}/L_\text{EN} (downlink limited). When Mβ†’NM \to N: NDT = 1 (baseline reached). When Cβ†’0C \to 0: NDT β†’βˆž\to \infty unless M=NM = N.

Proof
Fronthaul contribution

Each EN needs (1βˆ’ΞΌ)(1 - \mu) files' worth of content delivered from cloud (the rest is cached). Per-EN fronthaul time: K(1βˆ’ΞΌ)/(NENCF)K (1-\mu)/(N_\text{EN} C_F) per channel use.

Downlink contribution

The NENN_\text{EN} ENs cooperatively deliver KK files to KK users via MIMO BC with effective antenna count Leff=LENβ‹…NENL_\text{eff} = L_\text{EN} \cdot N_\text{EN}. Delivery time: K(1βˆ’ΞΌ)/(NENβ‹…min⁑(LEN,K))K(1 - \mu)/(N_\text{EN} \cdot \min(L_\text{EN}, K)).

Total NDT

NDT is the sum of fronthaul and downlink contributions, lower- bounded by 1 (baseline). The derivation in Sengupta-Tandon- Simeone '17 gives the precise bounds.

Saturation

At μ=1\mu = 1: fully cached; no fronthaul needed; NDT = 1. At μ=0\mu = 0, C=∞C = \infty: NDT = K/(NENLEN)K/(N_\text{EN} L_\text{EN}) (MU-MIMO baseline).

NDT vs Fronthaul Capacity

NDT Ξ”\Delta as a function of per-EN fronthaul capacity CC, for fixed cache size and varying memory ratios. At low CC, NDT is large (fronthaul bottleneck); it decreases as CC grows, hitting 1 when fronthaul is abundant. Higher cache ΞΌ\mu shifts the curve down β€” cache reduces the fronthaul demand.

Parameters
4
4
0.3

Example: NDT Computation for a 5G Small Cell

A small-cell C-RAN has NEN=2N_\text{EN} = 2 ENs, K=8K = 8 users, LEN=4L_\text{EN} = 4 antennas each, fronthaul C=2C = 2 files/use, cache M/N=0.25M/N = 0.25. Compute the NDT and identify the bottleneck.

Solution
Fronthaul contribution

K(1βˆ’ΞΌ)/(NENC)=8β‹…0.75/(2β‹…2)=1.5K(1-\mu)/(N_\text{EN} C) = 8 \cdot 0.75/(2 \cdot 2) = 1.5.

Downlink contribution

K(1βˆ’ΞΌ)/(NENβ‹…LEN)=8β‹…0.75/(2β‹…4)=0.75K(1-\mu)/(N_\text{EN} \cdot L_\text{EN}) = 8 \cdot 0.75/(2 \cdot 4) = 0.75.

Total NDT

Ξ”=max⁑(1,1.5+0.75)=2.25\Delta = \max(1, 1.5 + 0.75) = 2.25. Above the baseline of 1.

Bottleneck

Fronthaul (1.5) > downlink (0.75). Increasing fronthaul CC from 2 to 4 would halve the fronthaul contribution to 0.75, giving Ξ”=1.5\Delta = 1.5. Doubling cache ΞΌ\mu to 0.5 gives fronthaul =8β‹…0.5/(2β‹…2)=1.0= 8 \cdot 0.5/(2 \cdot 2) = 1.0, downlink =0.5= 0.5, total Ξ”=1.5\Delta = 1.5. Equivalent improvement via either resource.

Operational insight

At this operating point, cache and fronthaul are substitutes: doubling either has the same effect on NDT. At other operating points the substitution rate changes; the NDT surface captures this.

Key Takeaway

NDT unifies cache size and fronthaul capacity into a single latency metric. It removes the SNR dependence and exposes the architectural tradeoff directly. NDT = 1 is the ideal (infinite fronthaul baseline); higher NDT means slower delivery. System design with the NDT framework: pick (M,CF)(M, C_F) to meet a latency budget.

πŸŽ“CommIT Contribution(2017)

The NDT Framework for Cache-Aided Cloud-RAN

A. Sengupta, R. Tandon, O. Simeone, G. Caire β€” IEEE Transactions on Information Theory

The NDT framework was introduced in a series of papers by Simeone's group, with Caire and collaborators co-authoring extensions to multi-antenna and cooperative settings. The key contribution:

  1. NDT as a unifying metric. Captures the cache-fronthaul tradeoff in a scale-invariant number that removes SNR dependencies.
  2. Achievability schemes. Cooperative Lampiris-Caire style delivery with fronthaul-aware placement; time-share between cache-heavy and fronthaul-heavy modes.
  3. Converse bounds. Information-theoretic lower bounds on NDT, tight at several operating points.
  4. Practical implications. The framework informs the architectural choice of where to put cache and how much fronthaul to provision.

The CommIT follow-up work has extended NDT to mixed-traffic (Park-Caire 2020), privacy (Wan-Caire 2022), and massive MIMO (Lampiris-Bhattacharjee-Caire 2023). Chapter 8 of this book presents the baseline framework; subsequent chapters touch on extensions.

cloud-ranndtcommitView Paper β†’

Historical Note: From C-RAN to Cache-Aided C-RAN

2011–2022

The C-RAN concept (China Mobile, 2011) preceded coded caching by three years. The initial motivation was centralizing baseband processing to reduce BS cost and enable coordinated multi-point transmission (CoMP). Caching at the edge was not part of the original vision.

Simeone and collaborators (2015–2017) connected the two: once you have ENs with local storage (needed for CoMP buffering anyway), why not pre-cache popular content? The NDT framework crystallized this intuition. By 2020, 3GPP Rel-16 added caching hooks at the DU/ CU levels. Cache-aided C-RAN is now a mainstream research topic with direct deployment paths.

The CommIT group's contribution: unifying the NDT framework with the multi-antenna coded caching theory of Lampiris-Caire. Fog massive MIMO β€” CommIT's flagship 6G architecture β€” is C-RAN + caching + massive MIMO, analyzed via NDT + DoF metrics.

The Cloud-RAN Architecture with Edge CachingThe Cache-Fronthaul Tradeoff