Source-linked AI summary
Impact of traffic mix on caching performance in a content-centric network
Christine Fricker, Philippe Robert, James Roberts, Nada Sbihi
TL;DR
The paper examines how a realistic mix of web, file sharing, user generated content, and video on demand affects bandwidth-memory tradeoffs in a two-layer cache hierarchy. It models the differing traffic, population, size, and popularity characteristics of these content types, then evaluates cache hit rates using the Che approximation for LRU performance. Results favor caching VoD in access-router caches, while the other content types require large core storage capacities for significant bandwidth reduction.
Problem
The paper asks how caching can achieve favorable memory-bandwidth tradeoffs for the diverse content types dominating Internet traffic.
Method
The paper evaluates a two-layer cache hierarchy under a realistic four-type traffic model and applies the Che approximation to estimate LRU hit rates across layers.
Results
VoD is favorably cached in layer 1 access-router stores, whereas significantly reducing bandwidth for web, file sharing, and UGC requires a large cache of at least 100 TB in the core.
Takeaways & Limitations
Layer 1 caches may preferably be dedicated to VoD, while the other three content types are handled in very large-capacity layer 2 storage.
Takeaways & Limitations
The analysis relies on estimated content characteristics, and VoD performance depends critically on establishing its popularity law more reliably.
Abstract
from arXiv · showhide
For a realistic traffic mix, we evaluate the hit rates attained in a two-layer cache hierarchy designed to reduce Internet bandwidth requirements. The model identifies four main types of content, web, file sharing, user generated content and video on demand, distinguished in terms of their traffic shares, their population and object sizes and their popularity distributions. Results demonstrate that caching VoD in access routers offers a highly favorable bandwidth memory tradeoff but that the other types of content would likely be more efficiently handled in very large capacity storage devices in the core. Evaluations are based on a simple approximation for LRU cache performance that proves highly accurate in relevant configurations.
1. Introduction
The paper examines how caching can reduce bandwidth in content-centric networks by evaluating a two-layer hierarchy under realistic Internet traffic. It models distinct content types and uses established cache-replacement assumptions and an LRU approximation to assess performance.
- 96% of Internet traffic involves users retrieving content, motivating analysis of the bandwidth-memory tradeoff from caching.
- The modeled hierarchy places many similarly sized caches in access routers and a coordinated large cache in the network core.Requests missing from layer 1 are routed to layer 2.
- The evaluation measures bandwidth savings between cache layers and reductions in traffic arriving from sources, peers, and transit providers.
- Requests follow the independent reference model, with LRU as the main replacement policy and LFU considered for potential hit-rate gains.
- The demand model distinguishes web, file sharing, UGC, and VoD by traffic proportions, object and population sizes, and popularity distributions.
- The study adapts the Che approximation to realistic mixed traffic and reports high accuracy across broader populations and popularity distributions than previously considered.
2. Internet content characteristics
The paper characterizes four Internet content classes using estimated traffic, population, object-size, and popularity properties. These estimates motivate contrasting Zipf assumptions for the content classes, especially the less-characterized VoD category.
- 96% of Internet traffic is classified as content retrieval, divided here into web, file sharing, UGC, and VoD.The traffic mix is based on Cisco Visual Networking Index estimates and forecasts.
- Web content is conservatively modeled as 10^11 named elements with a mean size of 10 KB.
- File-sharing estimates use Demonoid statistics and a representative torrent sample with a mean object size of 7.4 GB.
- UGC is estimated at 5×10^8 YouTube videos averaging 10 MB, while VoD objects average around 100 MB in much smaller catalogues.
- The first three content types each total about 1 petabyte, three orders of magnitude above the assumed 1-terabyte VoD volume.
- Popularity distributions: Measured torrent popularity closely matches Zipf(0.823) on Demonoid and Zipf(0.75) on Pirate Bay.
- Popularity distributions: UGC popularity estimates range from Zipf exponents of 0.56 to approximately 0.88, while VoD may have a two-component distribution or Zipf exponents between 0.65 and 1.
- Popularity distributions: For evaluation, web, file sharing, and UGC use Zipf(0.8), while VoD is tested with contrasting Zipf(0.8) and Zipf(1.2) laws.
3. Caching homogeneous content
For homogeneous traffic, cache hit rates depend strongly on popularity concentration and cache organization. The Che approximation accurately estimates LRU performance and supports analysis of two-layer hierarchies.
- 3.1. LFU performance.: LFU maximizes hit rate under the independent reference model and provides a benchmark for evaluating LRU.The model assumes object requests follow independent Poisson processes, with request probability determined only by object popularity.
- 3.1. LFU performance.: Caching is significantly less effective when the Zipf parameter α is below 1, and population size materially affects performance.For α < 1, hit rate versus normalized cache size tends toward a limit as population grows; for α > 1, hit rate versus absolute cache size tends toward a limit.
- 3.2. LRU performance.: The Che approximation estimates each LRU object's hit rate as 1 − e^(-q(n)T_C), with T_C solved from the expected cache occupancy.The approximation treats T_C as the time needed for C distinct objects to be requested and is accurate across varied populations and popularity distributions.
- 3.2. LRU performance.: Simulation and Che-approximation curves agree for both a 10^4-object Zipf(.8) population and a 16-object geometric population.The approximation is also exact for a uniform popularity distribution.
- 3.2. LRU performance.: LRU is less efficient than LFU, especially for caches up to roughly 10% of a 10^4-object population under Zipf(.8) popularity.Figure 5 compares the two policies for Zipf(.8) and Zipf(1.2) distributions.
- 3.3. A two-layer hierarchy.: In a two-layer hierarchy, the second layer receives residual popularity q′(n) = q(n)(1 − h(n)), enabling hit-rate estimation by reapplying the Che approximation.The overall hit rate is calculated from the layer-1 and layer-2 hit rates while accounting for coordinated overflow traffic.
4. Caching the content mix
The realistic mixed-content model evaluates LFU and LRU caching across a two-layer hierarchy, showing sharply different cache behavior among content types. VoD is effective in small access-router caches, while other content types require much larger core storage for comparable hit rates.
- Mixed-content model: The cache evaluates four content types together, applying LFU or LRU replacement without discriminating by type and assuming equal object sizes within each type.The model represents web, file sharing, UGC, and VoD through differing traffic proportions, populations, object sizes, and popularity distributions.
- LFU performance: The LFU formulation parameterizes cache size and hit rate by storing objects whose arrival rates exceed a threshold.Relations (2) and (3) define a parametric hit-rate-versus-cache-size plot.
- LRU performance: The mixed-content LRU approximation assigns each object hit rate 1 − e^−q_i(n)T_C, with T_C determined by the cache-size equation.The overall hit rate is obtained by summing request rates multiplied by the corresponding object hit rates.
- Overall results: LRU has lower hit rates than LFU, especially for small to medium caches and Zipf(.8), while VoD popularity most affects caches where VoD occupies a substantial share.The adapted Che approximation is reported as highly accurate over the evaluated configurations.
- Two-layer hierarchy: VoD reaches a high hit rate with roughly 1 TB in layer 1 and little layer 2 usage, whereas UGC needs around 100 TB across both layers for a similar hit rate.UGC, file sharing, and web exhibit roughly similar behavior; the VoD Zipf exponent strongly affects VoD but not UGC hit rates.
- Bandwidth-saving comparison: Dedicating the 1 TB layer-1 cache to VoD is significantly more efficient under Zipf(.8), but provides little to no gain under Zipf(1.2).With accentuated VoD popularity, the layer-1 cache could instead be reduced.
5. Conclusion
The two-layer cache hierarchy has different implications across content types: the first three require very large core storage, while VoD is better suited to access-router caches. These conclusions depend on estimated traffic characteristics, especially VoD popularity, although the Che approximation was numerically validated as accurate for the study’s relevant configurations.
- 100 TB or more is needed to significantly reduce bandwidth requirements for web, file sharing, and UGC, each modeled at around 1 petabyte with Zipf exponent 0.8.Such capacity would likely be economically provided in the network core, or layer 2.
- VoD’s 1 TB catalogue and high, increasing traffic share make access-router caching a favorable bandwidth-memory tradeoff.Hit-rate comparisons suggest dedicating layer 1 caches to VoD rather than caching all content indiscriminately.
- The analysis relies on estimated content characteristics that should be made more reliable, particularly VoD’s popularity law because performance depends critically on it.
- The Che approximation was numerically validated as highly accurate, especially for the large populations and cache sizes relevant to this study.The mathematical arguments explaining this precision remain to be established.