Data Acquisition for Real-time Decision-making under Freshness Constraints Shaohan Hu , Shuochao Yao, Haiming Jin, Yiran Zhao, Yitao Hu, Xiaochen Liu, Nooreddin Naghibolhosseini, Shen Li, Akash Kapoor, William Dron, Lu Su, Amotz Bar-Noy, Pedro Szekely, Ramesh Govindan, Reginald Hobbs, Tarek Abdelzaher
2
Chaotic, dynamic environments In response, need to decide… what course of action to take how to carry it out 3
Shelter has Comm. vacancy center up Road not Disease DECISION blocked controlled Medic team Food store sta ff ed not flooded 4
DECISION OR AND AND Comm. Disease Food store Shelter has Road not Medic team center up controlled not flooded vacancy blocked sta ff ed 5
Decision Maker DECISION 6
7
Resource Limitation Environment Dynamics 8
1 9
10
OR AND AND 11
12
Order? 13
OR AND AND D E A B C 14
Intuition OR Maximize probability of short-circuiting per unit cost AND AND First examine the course of action that’s most likely to succeed Within a course of action, first examine the D E A B C condition that’s most likely to fail 15
OR AND AND D E A B C C 4 3 P ( True ) 0.3 0.6 R (1-0.3)/4 (1-0.6)/3 =0.175 =0.133 16
OR C (D->E) = 4+0.3(3) = 4.9 AND AND C (E->D) = 3+0.6(4) = 5.4 D E B C C 4 3 P ( True ) 0.3 0.6 R (1-0.3)/4 (1-0.6)/3 =0.175 =0.133 17
OR AND AND 7 4.9 C A B A B C D E P ( True ) 0.1 0.18 R 0.1/7=0.014 0.18/4.9=0.037 18
C (L->R) = 7+(1-0.1)x4.9 = 11.41 C (R->L) = 4.9+(1-0.18)x7 = 10.64 OR AND AND 7 4.9 C A B A B C D E 0.1 0.18 P ( True ) R 0.1/7=0.014 0.18/4.9=0.037 19
key: Short-circuit probability cost 20
2 21
OR AND AND 23
24
retrieval latency { { freshness interval 25
26
Some random ordering 27
expired Some random ordering 28
expired bad resolution Some random ordering 29
30
Latest Deadline First (LDF) 31
good resolution Latest Deadline First (LDF) 32
LDF - Latest Deadline First Inspired by EDF: data objects with later freshness deadlines are retrieved sooner Optimal: if LDF cannot avoid freshness deadline violation, no sequential order can 33
LDF , what about Cost? 34
LDF, what about Cost? 35
LDF, what about Cost? LDF satisfies data freshness constraints 36
LDF, what about Cost? This might have lower expected cost 37
LDF, what about Cost? This might have lower expected cost Greedily rearrange LDF order to reduce the expected data retrieval cost 38
LDF , still Freshness Violations? 39
LDF, still Freshness Violations? 40
LDF, still Freshness Violations? 41
LDF, still Freshness Violations? 42
LDF, still Freshness Violations? expired 43
LDF, still Freshness Violations? parallel retrieval 44
LDF, still Freshness Violations? good resolution 45
LDF, still Freshness Violations? good resolution Increment parallel retrieval level until freshness constraints are met 46
vLDF Data Retrieval Compute LDF order Greedily rearrange LDF order to reduce the expected data retrieval cost Gradually increment parallel retrieval level until freshness constraints are met 47
Evaluation Simulation experiments An application scenario 48
Simulation Experiments Baselines LCF - Lowest Cost Source First SCB - Shortcircuit Benefit only PbP - Probability based Prediction Settings % fast changing data: 40~100%, default 70% # Action size: 4~10, default 6 Data object size: 3~5 MB, default 3.45 MB Network bandwidth: 3.5~6.5 KBps, default 5 KBps Transmission latency fluctuation: -3~3 min, default 0 49
Simulation Results 150 100 Request Resolution Rate (%) vLDF Retrieval Cost Ratio (%) SCB 80 PbP 100 LCF vLDF 60 SCB PbP 40 LCF 50 20 0 0 40 50 60 70 80 90 100 40 50 60 70 80 90 100 Ratio of Fast − Changing Data Items (%) Ratio of Fast − Changing Data Items (%) Varying % of fast changing data 50
Simulation Results 200 100 Request Resolution Rate (%) vLDF Retrieval Cost Ratio (%) SCB 80 150 PbP LCF vLDF 60 SCB 100 PbP 40 LCF 50 20 0 0 4 5 6 7 8 9 10 4 5 6 7 8 9 10 Number of Conditions per Action Number of Conditions per Action Varying action size 51
Simulation Results 150 100 Request Resolution Rate (%) vLDF Retrieval Cost Ratio (%) SCB 80 PbP 100 LCF vLDF 60 SCB PbP 40 LCF 50 20 0 0 2000 2500 3000 3500 4000 4500 5000 2000 2500 3000 3500 4000 4500 5000 Average Data Item Size (KB) Average Data Item Size (KB) Varying data object size 52
Simulation Results 150 100 Request Resolution Rate (%) vLDF Retrieval Cost Ratio (%) SCB 80 PbP 100 LCF vLDF 60 SCB PbP 40 LCF 50 20 0 0 3.5 4 4.5 5 5.5 6 6.5 3.5 4 4.5 5 5.5 6 6.5 Network Bandwidth (KBps) Network Bandwidth (KBps) Varying network bandwidth 53
Simulation Results 150 100 Request Resolution Rate (%) vLDF Retrieval Cost Ratio (%) SCB 80 PbP 100 LCF vLDF 60 SCB PbP 40 LCF 50 20 0 0 − 3 − 2 − 1 0 1 2 3 − 3 − 2 − 1 0 1 2 3 Network Transmission Delay Fluctuation − Mean Network Transmission Delay Fluctuation − Mean Varying network transmission fluctuation 54
Application: Route Finding Find routes for <src, dst> pairs Each candidate route: AND of its segments Routing result: OR of all candidate routes Visual verification for route segment conditions 55
Results of 5 Runs vLDF Cost (KB) PbP Cost (KB) vLDF Time (s) PbP Time (s) 516 685 164 255 343 598 150 206 319 485 160 248 506 1093 165 372 524 1042 175 206
Conclusion Environment dynamics & resource limitations affect real-time decision-making Efficient data acquisition algorithm Promising results through simulations and concrete route finding application scenario 59
Thanks
Recommend
More recommend
