Ludicrous speed

Fantastic stories about performance requirements

Transcript

1. fantastic stories about performance requirements Ludicrous Speed
7. PANIC!
8. Modern human experimentations
11. Big business impact Big response delay
12. No silver bullets today
13. Perception of time What happens on internets? What is “OK” response time? What is “OK” load? Summary PLAN
14. Perception of time What happens on internets? What is “OK” response time? What is “OK” load? Summary PLAN
15. Perception of time What happens on internets? What is “OK” response time? What is “OK” load? Summary PLAN
16. Perception of time What happens on internets? What is “OK” response time? What is “OK” load? Summary PLAN
17. Perception of time What happens on internets? What is “OK” response time? What is “OK” load? Summary PLAN
18. Perception Of Time
19. 100-200 ms 500-1000 ms 2-5 seconds 7-10 seconds instantaneous reaction uninterrupted reaction some reaction oof…
20. 100-200 ms 500-1000 ms 2-5 seconds 7-10 seconds instantaneous reaction uninterrupted reaction some reaction oof…
21. 100-200 ms 500-1000 ms 2-5 seconds 7-10 seconds instantaneous reaction uninterrupted reaction some reaction oof…
22. 100-200 ms 500-1000 ms 2-5 seconds 7-10 seconds instantaneous reaction uninterrupted reaction some reaction oof…
23. 100-200 ms 500-1000 ms 2-5 seconds 7-10 seconds instantaneous reaction uninterrupted reaction some reaction oof… *20 years old data
24. Modern data?
25. 3… may be 5 seconds max
26. 3… may be 5 seconds max Even faster in future
27. Weber–Fechner law 1 2
28. Weber–Fechner law 1 2
29. Weber–Fechner law 1 2
30. Weber–Fechner law 1 2
31. Weber–Fechner law 1 2 just noticeable difference (jnd)
32. Weber–Fechner law just noticeable difference (jnd) For time lengths of up to 30s - jnd is somewhere between 7% and 18%
33. Weber–Fechner law Faster Slower
34. Weber–Fechner law 20s Faster Slower
35. Weber–Fechner law 20s 24s 16s -20% +20% Faster Slower
36. “Slightly” faster
37. Perception of average between t1 and t2 is: sqrt(t1*t2)
38. Example
39. Benchmark against 30s 60s less than 36s less than 42s less than 24s slow same slightly slower faster
40. Benchmark against 30s 60s less than 36s less than 42s less than 24s slow same slightly slower faster
41. Benchmark against 30s 60s less than 36s less than 42s less than 24s slow same slightly slower faster
42. Benchmark against 30s 60s less than 36s less than 42s less than 24s slow same slightly slower definitely faster
43. Is it worth to optimise? Target performance Unnoticeable degradation
44. Is it worth to optimise? Target performance Unnoticeable degradation
45. Is it worth to optimise? Target performance Unnoticeable degradation
46. Why so serious?
47. Why so serious?
48. What Happens on Internets?
51. Response time: DNS request Establish TCP connection Generate request to server Send request to server Server generates response Server sends response Page rendering on browser
52. Response time: DNS request Establish TCP connection Generate request to server Send request to server Server generates response Server sends response Page rendering on browser
53. Response time: DNS request Establish TCP connection Generate request to server Send request to server Server generates response Server sends response Page rendering on browser
54. Response time: DNS request Establish TCP connection Generate request to server Send request to server Server generates response Server sends response Page rendering on browser
55. Response time: DNS request Establish TCP connection Generate request to server Send request to server Server generates response Server sends response Page rendering on browser
56. Response time: DNS request Establish TCP connection Generate request to server Send request to server Server generates response Server sends response Page rendering on browser
57. Response time: DNS request Establish TCP connection Generate request to server Send request to server Server generates response Server sends response Page rendering in a browser
58. Response time: DNS request Establish TCP connection Generate request to server Send request to server Server generates response Server sends response Page rendering in a browser
59. Fun facts! Min time to pass 1/2 of globe: 67ms
60. Later ;) Server generates response
61. Let’s talk a bit about that Page rendering in a browser
62. In a complex projects 60% of time could be easily spent on rendering
63. There’s space to optimise
64. There’s space to optimise
65. There’s space to optimise
66. Time budget: Load and render critical parts first
67. What Is “OK” Response?
68. It depends
69. In general case response longer than 300ms means nobody cares
70. In general case response longer than 300ms means nobody cares Usually it’s more than 1 second on a client
71. It should be fast
72. It should be fast Bad requirement
73. It should be faster than 300ms
74. It should be faster than 300ms Bad requirement
75. Long tail
76. 95% < 300ms 99% < 500ms
77. 95% < 300ms 99% < 500ms
78. 95% < 300ms 99% < 500ms OK requirement
79. 95% < 300ms 99% < 500ms OK requirement But still depends
80. Remember about 20% difference!
81. Remember to measure competitors!
82. Oof… Done with time!
83. Oof… Done with time! There are 1 more to go
84. What Is “OK” Load?
85. What measurement units to use?
86. What is “average user”?
87. “Average human” have less than 1 arm
88. Requests Per Second (RPS) is what server understands
89. How much RPS we should be able to handle?
90. EXTREMELY HARD QUESTION
91. Pulse of the Web
92. Spotify Average Monday in Stockholm
93. Spotify Specific Monday in Stockholm
98. Spotify Average Monday in Eastern US
99. Financial Times weekly traffic changes
100. Financial Times weekly traffic changes Extremely different for different clients
101. Financial Times weekly traffic changes Extreme daily features
102. Thanksgiving in US Events could make a big difference
103. Bombing in Moscow subway 29 march 2010
104. Bombing in Moscow subway 29 march 2010 First explosion 7:57 Load growth 8:10
105. Think like a user Look for daily/weakly/yearly cycles
106. Think like a user Look for daily/weakly/yearly cycles
107. Event based service? Prepare for extremely huge spikes
108. Do closed beta And watch for load profile
109. It’s growing
110. Very Short Summary
111. Find out what is OK response time for you Remember about long tail Find out possible max load Test your assumptions
112. Find out what is OK response time for you Remember about long tail Find out possible max load Test your assumptions
113. Find out what is OK response time for you Remember about long tail Find out possible max load Test your assumptions
114. Find out what is OK response time for you Remember about long tail Find out possible max load Test your assumptions
115. APPLAUSE
116. QUESTIONS?
117. Magic Numbers Some
118. 4 datacenters More than 5K servers ~50 millions of unique users More than 100 people working on backend infrastructure and operations
119. DATACENTERS: Ashburn San Jose Stockholm London
120. QUESTIONS?
121. Sergey Vysotskiy vsp@spotify.com