Designing for Performance. Martin Thompson

Transcription

1 Designing for Performance Martin Thompson

2

3

4 Feynman is becoming a real pain. He has the greatest scientific honesty of anyone I ve ever meet - William P Rogers

5 The impact of QED cannot be overestimated. It explains everything that is not explained by gravity. It s also the most accurate theory ever tested by experiments on Earth. - Freeman Dyson

6

7 It does not matter how intelligent you are, if you guess and that guess cannot be backed up by experimental evidence then it is still a guess. - Richard Feynman

8 How do we Design for Performance?

9 1. What is Performance? 2. What is Clean & Representative? 3. Implementing efficient Models 4. How to Performance Test

10 Performance

11 Throughput (aka Bandwidth)

12 Response Time (aka Latency)

13 Scalability

14

15 Response Time Queuing Theory Utilisation

16 Pro Tip: Ensure you have sufficient capacity

17 Can we go parallel to speedup?

18 Amdahl s Law time Sequential Process A B

19 Amdahl s Law time Sequential Process A B Parallel Process A A A A A B

20 Amdahl s Law time Sequential Process A B Parallel Process A A A A A B Parallel Process B A B B B B

21 Amdahl's Law

22 Universal Scalability Law C(N) = N / (1 + α(n 1) + ((β* N) * (N 1))) C = capacity or throughput N = number of processors α = contention penalty β = coherence penalty

23 Speedup Universal Scalability Law Processors Amdahl USL

24 Clean & Representative

25 - Clean Morally uncontaminated; pure; innocent - Oxford English Dictionary

26 - Representative Serving as a portrayal or symbol of something - Oxford English Dictionary

27 - Representative Code is the best place to capture current understanding of a model

28 Abstractions

29 Rules of Abstraction 1. Don t use abstraction 2. Don t use abstraction 3. Only consider abstracting when you see at least 3 things that ARE the same 4. Abstractions must pay for themselves 5. Beware DRY, the evil siren that tricks you into abstraction

30 Abstraction Megamorphism => Branch Hell

31 Abstraction Not Representative => Big Smell

32 Abstraction Say no to big frameworks!

33

34 Pro Tip: Abstract when you are sure of the benefits

35 Law of Leaky Abstractions All non-trivial abstractions, to some extent, are leaky. - Joel Spolsky

36 Law of Leaky Abstractions The detail of underlying complexity cannot be ignored.

37 the purpose of abstracting is not to be vague, but to create a new semantic level in which one can be absolutely precise - Dijkstra

38 How can we abstract memory systems?

39 - It s about 3 bets! 1. The Temporal Bet

40 - It s about 3 bets! 1. The Temporal Bet 2. The Spatial Bet

41 - It s about 3 bets! 1. The Temporal Bet 2. The Spatial Bet 3. The Striding Bet

42 Model Implementation

43 Coupling vs Cohesion

44 Coupling vs Cohesion public class Queue { private final Object[] buffer; private final int capacity; } // Rest of the code

45 Coupling vs Cohesion public class Queue { private final Object[] buffer; private final int capacity; } // Rest of the code

46 Coupling vs Cohesion Properties Bag

47 Pro Tip: Respect Locality of Reference

48 Relationships

49 Relationships Order Book Order

50 Relationships Order Book 1 Bids * 1 Offers * Order

51 Relationships Order Book Price Price 1 Bids * 1 Offers * Order

52 Relationships Order Book Price Price 1 Bids * {ordered, FIFO} 1 Offers * {ordered, FIFO} Order

53 Pro Tip: Make friends with your Data Structures

54 Pro Tip: Document, discuss, design tests, before going to code

55 Algorithms

56 Order of Algorithms

57 Order of Algorithms Magnitude of n?

58 Pro Tip: Know the cardinality of all significant relationships

59 Pro Tip: Algorithms are your key to service time

60 Batching

61 Amortise the expensive costs

62 Natural Batching Producers

63 Natural Batching << Amortise Expensive Costs >> Batcher Producers

64 Response Time Natural Batching Typical Possible Load

65 Pro Tip: Batch processing is not just for offline

66 Branches, branches, branches...

67 Branches public void dostuff(list<string> things) { if (null == things things.isempty()) { return; } } for (String thing : things) { // Do useful work }

68 Branches public void dostuff(list<string> things) { if (null == things things.isempty()) { return; } } for (String thing : things) { // Do useful work }

69 Branches public void dostuff(list<string> things) { for (String thing : things) { // Do useful work } }

70 Pro Tip: Respect the Principle of least surprise

71 Loops

72 Loops If I had more time, I would have written a shorter letter. - Blaise Pascal

73 Loops L µops

74 Loops L µops LBB 28 µops LBB 28 µops

75 Pro Tip: Craft major loops like good prose

76 Composition

77 Composition Size matters

78 Composition Inlining is THE optimisation. - Cliff Click

79 Composition Single Responsibility

80 Pro Tip: Small atoms can compose to build anything

81 Data

82 Data

83 Data

84 Data

85 Data

86 Data

87 Pro Tip: Embrace Set Theory and FP techniques

88 Performance Testing

89 Define Performance Goals

90 Establish Design Principles

91 Aeron Design Principles 1. Garbage free in steady state running 2. Smart Batching in the message path 3. Lock-free algos in the message path 4. Non-blocking IO in the message path 5. No exceptional cases in message path 6. Apply the Single Writer Principle 7. Prefer unshared state 8. Avoid unnecessary data copies

92 How to measure response time?

93 Response Time Histograms

94 Response Time Histograms Mode

95 Response Time Histograms Mode Median

96 Response Time Histograms Mode Median Mean

97 Coordinated Omission Source: Gil Tene (Azul Systems)

98 HdrHistogram

99 Java Microbenchmark Harness

100 CPU Performance Counters

101 Performance test as part of Continuous Integration

102 Can your acceptance tests run as performance tests?

103 Build telemetry into production systems

104 AGAIN!!! Build telemetry into production systems

105 Counters of: Queue Lengths Concurrent Users Exceptions Transactions - orders, trades Etc.

106 Histograms of: Response Times Service Times Queue Lengths Concurrent Users Etc.

107 In closing

108 Clean => Uncontaminated

109 Representative => True Portrayal

110 Does it pass the Out Loud test?

111 Measure Don t Guess!!!

112

113 Questions? It does not matter how intelligent you are, if you guess and that guess cannot be backed up by experimental evidence then it is still a guess. - Richard Feynman