Coupling Caching and Forwarding: Benefits, Analysis & Implementation

Transcription

1 Coupling Caching and Forwarding: Benefits, Analysis & Implementation Dario Rossi Giuseppe Rossini

2 Context Information centric neworks Mobility, naming, security,, and Caching Caching on Information centric networks Not a goal on its own, but a means (e.g., congestion-aware to reduce load on bottleneck links, or cost-aware to reduce ISPs opex) General viewpoint: Efficiently allocate/use caching resources Important notions Temporary vs Persistent On-path vs Off-path This talk B u1 u2 D R1 F A C E u3 How to efficiently find the closest,possibly off-path, content replica

3 Terminology ICN design space is wide We may use FIB toward custodian if available We ignore proactive/periodic dissemination of cached copies availability on control plane We focus on reactive discovery of cached copies on the data plane Algorithmically speaking, a cache network can be identified as a triple F is the Forwarding strategy D is the cache Decision policy R is the cache Replacement policy D=0 F=0 F=N D=1 C R C

4 Related work Typically, focus on either F or D, but not both (R=LRU as replacement policy makes sense) F=SPR by default Naive, on path Ideal, off path D=LCE by default

5 Agenda Benefits Interest of joint <F,D> Comparison with several CDN strategies [16] 26-1 Analysis Modeling of <inrr,lce,lru> Extend <SPR,LCE,LRU>, aka anet [31] 10x10 Implementation Notion of Meta-interest to avoid pollution across paths Spoiler: Arbitrarely close to inrr, slight delay tradeoff [16] S. K. Fayazbakhsh, Y. Lin, A. Tootoonchian, A. Ghodsi,T. Koponen, B. M. Maggs, K. Ng, V. Sekar, and S. Shenker. Less pain, most of the gain: Incrementally deployable. In ACM SIGCOMM, [31] E. J. Rosensweig, J. Kurose, and D. Towsley. Approximate Models for General Cache Networks. In IEEE INFOCOM, 2010

6 Coupling Caching and Forwarding: Benefits, Analysis & Implementation Coupling Caching and Forwarding: Benefits, Analysis & Implementation (scenario details in the paper) (scenario scripts available at) Part I

7 CDN vs ICN scenario ICN Scope CDN Scope

8 CDN Policies Edge Naive CDN, with all caches operating independently EdgeCoop Smarter CDN, all caches cooperate (with inrr) EdgeNormCoop As EdgeCoop, but individual caches have double size (network-wide same budget w.r.t. ICN scope)

9 ICN Policies <SPR,LCE> Naive ICN: on-path caching toward repository <NRR,LCE> Access to off-path copies, but cache pollution due to LCE <NRR,LCD> Smarter ICN: off-path copies, limits cache pollution via LCD Configurations: 2-levels vs Ubiquitous caches (as for EdgeNormCoop, keeping the same overall cache budget)

10 Average distance [hop] CDN vs ICN policies Edge 0 naive CDN Edge on-path ICN smart CDN naive off-path IC N smart CDN smart off-path ICN <SPR,LCE> 2 Levels Ubiquitous 0.2 EdgeCoop <inrr,lce> EdgeNormCoop <inrr,lcd> Redundancy probability µ <SPR,LCE> EdgeCoop <inrr,lce> EdgeNormCoop Levels Ubiquitous <inrr,lcd>

11 Average distance [hop] CDN vs ICN policies Edge 0 naive CDN Edge on-path ICN smart CDN naive off-path IC N smart CDN smart off-path ICN <SPR,LCE> 2 Levels Ubiquitous 0.2 EdgeCoop <inrr,lce> EdgeNormCoop <inrr,lcd> Redundancy probability µ <SPR,LCE> EdgeCoop <inrr,lce> EdgeNormCoop Levels Ubiquitous <inrr,lcd>

12 Average distance [hop] CDN vs ICN policies Edge naive CDN Edge Remark on-path#1 ICN smart CDN nnaaiviveeooffff-p-paathth NN Based on thisicicdifference [16] smart CDN dismisses caching, smart off-paubiquitous th ICN 0 arguing that most of the gain is 2 Levels achievable painlessly at the edge Ubiquitous 0.2 <SPR,LCE> EdgeCoop <inrr,lce> <inrr,lce> EdgeNormCoop <inrr,lcd> Redundancy probability µ <SPR,LCE> EdgeCoop <inrr,lce> EdgeNormCoop Levels Ubiquitous <inrr,lcd>

13 Average distance [hop] CDN vs ICN policies Edge naive CDN Edge Remark #2 on-path ICN smart CDN naive off-path N conservative Our resultsicare smart CDN as <inrr,lce> is better smartinoff[16] -path ICN than EdgeNormCoop 0 2 Levels Ubiquitous 0.2 <SPR,LCE> EdgeCoop <inrr,lce> EdgeNormCoop <inrr,lcd> Redundancy probability µ <SPR,LCE> EdgeCoop <inrr,lce> EdgeNormCoop Levels Ubiquitous <inrr,lcd>

14 Average distance [hop] CDN vs ICN policies Edge naive CDN Remark #3 on-path ICN 0 Edge Remark #4 <SPR,LCE> smart CDN EdgeCoop naive off-path CN appear Benefits ofiicn Ubiquitous only caching exhibits <inrr,lce> smart CDN awith incremental benefits (aggregation) EdgeNormCoop sm rt off-pmeta-caching ath ICN <inrr,lcd> with enough path diminishing diversity returns (popularity skew) 2 Levels Ubiquitous Redundancy probability µ <SPR,LCE> EdgeCoop <inrr,lce> EdgeNormCoop 2Levels Ubiquitous <inrr,lcd>

15 Average distance [hop] CDN vs ICN policies 4.4 Edge Remark #5 naive CDN 4.2 on-path ICN <SPR,LCE> 4 smart CDN Technico economic studies 3.8 EdgeCoop naive off-path IC N 3.6 needed to assess if benefits <inrr,lce> smart CDN 3.4 smart off-path EdgeNormCoop justify deployment I C N 3.2 <inrr,lcd> ICN success unlikely driven 3 2 Levels only by caching Ubiquitousbenefits Redundancy probability µ Edge <SPR,LCE> EdgeCoop <inrr,lce> EdgeNormCoop 1 2Levels Ubiquitous <inrr,lcd>

16 Coupling Caching and Forwarding: Benefits, Analysis & Implementation Coupling Caching and Forwarding: Benefits, Analysis & Implementation Part II

17 Modeling inrr The Ideal Nearest Replica Routing (inrr) [2] is nice and worth modeling Let start from the anet model [31] for SPR

18 Modeling inrr Similar to anet Define split ratio among multiple paths Applies split ratio to miss stream

19 Model accuracy anet suffers from IRM violation due to miss-stream far from the leafs LRU approximation very accurate for leaf nodes inrr shorter paths and larger fanout, make IRM violation effect less pronounced

20 Coupling Caching and Forwarding: Benefits, Analysis & Implementation Coupling Caching and Forwarding: Benefits, Analysis & Implementation Part III

21 Architecture General scheme: forward interests along either FIB for persistent on-path copies, proactively advertised by routing TFIB for temporary off-path cached copies, reactively discovered via exploration Explore Exploit time Interests Data TFIB={} FIB={x:3} TFIB={x:1} FIB={x:3} 1 3 TFIB={} FIB={x:3}

22 Meta-interests The Ideal Nearest Replica Routing (inrr) [2] is nice and worth implementing Explore via scoped flooding, at distance d hops 1-phase design (NRR ) Use regular interest packets i i i i i i All cache hits return real data. Redundant for popular content, Useless for unpopular content 2-phase design (NRR ) Use meta-interest packets i i i i Hits indicate Afterward, actual requests only sent through a single TFIB path

23 Practical NRR very close to inrr 10x10Grid 26-1Tree 1-phase Regular interests yield to cache pollution penalty LCE LCD Meta-interests + LCD meta-caching = <NRR,LCD> arbitrarily close to inrr 2-phases Meta-interests confine pollution to a single path Meta-caching limit pollution along the path

24 Implementing NRR Implementing ideal Nearest Replica Routing (inrr) Explore via scoped flooding, max distance d hops 1-phase design (NRR ) Use regular interest packets 2-phase design (NRR ) Use meta-interest packets Regular interest generate cache decision, triggering cache replacement and pollution Meta-interest return only data-availability indication, not data itself. Avoid Pollution! Fastest propagation, time to first chunk is lowest Slower propagation, time to first chunk longer (subsequent chunks are found at NRR)

25 Implementing NRR Implementing ideal Nearest Replica Routing (inrr) Explore via scoped flooding, max distance d hops 2-phase design (NRR ) In practice, delay penalty 1-phase design (NRR ) Use meta-interest packets Use regular interest packets only affects the first chunk, upper-bounded by network RTT Delay not necessarily longer Meta-interest return only Regular interest generate cache as content is closer, 1st chunk delay may even be shorter data-availability indication, decision, triggering cache not data itself. Avoid Pollution! replacement and pollution Fastest propagation, time to first chunk is lowest Slower propagation, time to first chunk longer (subsequent chunks are found at NRR)

26 Coupling Caching and Forwarding: Benefits, Analysis & Implementation Coupling Caching and Forwarding: Benefits, Analysis & Implementation Aftermath

27 Conclusions Contributions Benefits, analysis and implementation of joint forwarding and caching Meta-caching: implicit coordination of caching decisions along a single path Meta-interests: implicit coordination of caching decisions across multiple paths Limits Can we do better (e.g., 2-LRU instead of LCD)? How far are is <inrr,lcd,lru> from optimum? Only LCE is modeled Only synthetic topologies are simulated, with stationary popularity and no spatial skew

28 Implications Comparison guidelines Algorithmic suggestion for IRTF ICNRG baseline comparison :) Good practice = include <inrr,lcd,lru> (or better) :( Bad practice = using only naive <SPR,LCE,LRU> Comparison is feasible NRR implementation available in existing tools (e.g., ccnsim) inrr (and NRR ) simple to implement in other simulators Results consistent across simulators: Come see the demo!

29 ?? //

30 Backup slides

31 Interest of joint <F,D,(R=LRU)> Some gain Larger gain <SPR,LCE> <inrr,lce> <SPR,LCE> <inrr,lcd>

32 More realistic catalog/cache scales