Algorithms for and against the Cloud

Transcription

1 ETH Zurich Distributed Computing Group Algorithms for and against the Cloud Roger Wattenhofer

2 Disclaimer OSDI SenSys AAAI HotNets STOC SPAA FOCS PODC ICALP Mobicom SIGCOMM SODA EC

3 Algorithms for the Cloud

4 Algorithms for the Cloud Infrastructure

5 Algorithms for the Cloud just perfect

6 Algorithms for the Cloud Infrastructure

7

8

9

10

11

12 Find balanced separator of minimum size K.

13 Find balanced separator of minimum size K.

14 Find balanced separator of minimum size K.

15 Find balanced separator of minimum size K. Our result: almost linear time algorithm for small K. [Brandt, W., 2017]

16 [Brandt, W., 2017] Find balanced separator of minimum size K. Our result: almost linear time algorithm for small K. in a boring way

17 Algorithms for the Cloud just perfect

18 GPS for the Cloud

19 Just record 1ms of raw data

20 Coarse Time Navigation Exhaustive Search in Area

21 Also Robust to GPS Spoofing

22 Algorithms for the Cloud just perfect

23 $100B Revenue ¾ Online

24 Online Two Player Games Match Players Fast Waiting is Boooooring Match Players Well Similar Rating, Location, etc.

25 Min-Cost Perfect Matching With Delays (MPMD)

26 MPMD Example rating (space) time

27 MPMD Example rating (space) time

28 MPMD Example rating (space) time

29 MPMD Example rating (space) time

30 MPMD Example rating (space) space cost time cost time

31 MPMD Example rating (space) space cost time cost time

32 MPMD Example rating (space) space cost time cost time

33 MPMD Example rating (space) space cost time cost time

34 Haste Makes Waste!

35 MPMD Example rating (space) space cost time cost time

36 MPMD Example rating (space) time

37 MPMD Example rating (space) time

38 MPMD Example rating (space) algorithm cost optimal cost time

39

40 [Wang et al., 2018]

41 The O(log n) Algorithm

42 Approximate Metric by Tree Height = O(log n) E[Distortion] = O(log n) w Leaves = Nodes in Metric Space [Fakcharoenphol, Rao, Talwar 2004], [Bansal, Buchbinder, Gupta, Naor 2015]

43 Algorithm

44 Algorithm

45 = w Algorithm

46 Algorithm

47 Algorithm

48 Algorithm

49 Algorithm

50 Algorithm

51 Algorithm

52 Algorithm

53 Proof

54 Proof

55 Proof

56 Proof Total space cost = σ

57 Proof

58 Proof

59 Proof For each pair at least one timer running Total time cost 2 σ

60 Total Algorithm Cost = O(σ )

61 What about OPT?

62 Proof

63 Proof ALG OPT time

64 Proof ALG OPT time

65 Proof ALG OPT time

66 Proof ALG OPT time or ALG OPT time

67 Proof ALG OPT time or ALG OPT time cost = cost

68 Done?

69 Just One Little Thing

70 Proof ALG OPT time

71 Proof ALG OPT time

72 Proof ALG OPT time

73 Proof ALG OPT time

74 Proof ALG OPT time

75 Proof ALG OPT time

76 Proof ALG OPT time

77 Proof ALG OPT time

78 Proof ALG OPT time

79 OPT has an easy time

80 but only every other phase!

81 Total OPT Cost = Ω(σ )

82 Where is the log n coming from? Height = O(log n) for time E[Distortion] = O(log n) for space

83

84 Algorithms against the Cloud

85

86

87 2008

88 Blockchain

89

90 Blockchain Basics

91 Transaction

92 Transaction

93 Transaction

94 Transaction

95 Block

96 Blockchain

97 Blockchain is Replicated

98 Blockchain Distributed Systems & Cryptography (1982) (1976)

99 Blockchain Distributed Systems & Cryptography Fault-Tolerance & Digital Signatures

100 Rule of Thumb Blockchains* may disrupt your business if you use signatures. *or blockchain-like tech

101 Blockchain Variants

102

103 Permissionless / Open

104 Permissioned / Closed

105 Multiple Participants? yes no No Blockchain (use DB/Cloud) Writers Known? no Permissionless Blockchain yes Permissioned Blockchain

106 Privacy The Seven Blockchain Dimensions Persistence Energy

107 Blockchain Persistence Database Immutable Provable Fault-Tolerance Correct Crash Byzantine

108 Blockchain Speed 1 hour 1 minute 1 second Throughput 10 tx/s 10k tx/s 10m tx/s

109 Scalability 10 nodes 100 nodes 1000 nodes Blockchain

110 Energy Consumption

111

112 Economic Incentives Market / Energy Value 12 GW $1M/h $0.08/kWh

113 Proof of Work Hashrate Energy/Hash 1.3 GW GH/s 0.1 J/GH

114 Privacy The Seven Blockchain Dimensions Persistence Energy

115 What About Privacy?

116 It s Complicated.

117 Privacy Anonymity/Public Identity/Private

118 Research Issues Privacy Persistence Energy Solution to many problems: Layer 2 Plus: crypto, language (smart contracts), game theory, measurements,

119 emoney

120 Permissioned Blockchain & Payment Network

121 Permissioned Blockchain

122 Payment Network

123 Bitcoin Anonymity Open/Anarchic Blockchain Eventual Consistency Proof-of-Work emoney Accountability Closed/Private Paxos, PBFT, Strong Consistency Central Banks

124 evoting

125 What s Wrong with Paper?

126 Cost

127 Verifiability

128 Anonymity Identity Swapper Identity Mixer

129 Election Help

130 Democracy Beyond Yes or No 2025 Wie viel sollen die SRG-Gebühren pro Jahr kosten? 42.-

131 Don t bring a Blockchain to a Gunfight

132 So what s new, really?

133 Hello World! timing crashes omission Now solve consensus Byzantine Classical Adversary 137

134 Здравствуйте! meltdown spectre re-entrancy Now hold an election rowhammer Modern Adversary 138

135 Hype and Criticism First practical solution to a longstanding problem in computer science, Byzantine Generals. Satoshi solved a problem that academic computer scientists thought was impossible Bitcoin is digital gold, it will put us back onto a sound monetary policy Bitcoin will end wars A non-deliberate Ponzi scheme It s yet another eventually consistent database Flawed technology, inherently limited in scale and performance Unlikely to impact the finance sector

136 Would you rather fight? Cloud vs. Blockchain

137 Would you rather trust? big corporation relay hodler miner developer user Cloud vs. Blockchain

138 Thanks to lots of hardware Cloud vs. Blockchain

139 We at big corp will run your blockchain in our cloud!

140 What s this?

141 A Blockchain?

142 A Cloud?

143 A Distributed System!

144 Summary

145 Thank You! Questions & Comments? Thanks to my co-authors Vertex Separators: Sebastian Brandt Online With Delay: Yuval Emek, Shay Kutten Cloud GPS: Manuel Eichelberger

146 Abstract: Algorithms interact in two main ways with the cloud. There exist algorithms which are tailored for the cloud, for which the cloud is the perfect environment. Moreover, the cloud may also benefit from optimization algorithms, algorithms that make the cloud more efficient. The AlgoCloud program features papers which roughly fit one of the two, and I will also give a few examples in the first part of my talk. Apart from these algorithms for the cloud, I will also talk about algorithms against the cloud. Recently, blockchains are hyped to be a cloud competitor, sometimes even a cloud killer. In the second part of my talk we will discuss whether there is some truth to whether blockchains are going to threaten the successful cloud paradigm.