Tradeoffs, Bad Science, and Polar Bears – The World of Java Optimisation

by Holly Cummins

Slide 1

Slide 1 text

tradeoffs, bad science, and polar bears: the world of java optimisation www.cloud-builders.tech Holly Cummins Red Hat

Slide 2

Slide 2 text

#Quarkus @holly_cummins why optimise?

Slide 3

Slide 3 text

#Quarkus @holly_cummins why optimise?

Slide 4

Slide 4 text

#Quarkus @holly_cummins why optimise? 0.5s extra search page time

Slide 5

Slide 5 text

#Quarkus @holly_cummins why optimise? 0.5s extra search page time 20% drop in traffic

Slide 6

Slide 6 text

#Quarkus @holly_cummins why optimise? 0.5s extra search page time 20% drop in traffic 100 ms latency on page load

Slide 7

Slide 7 text

#Quarkus @holly_cummins why optimise? 0.5s extra search page time 20% drop in traffic 100 ms latency on page load 7% lower conversion rate

Slide 8

Slide 8 text

#Quarkus @holly_cummins why optimise? 0.5s extra search page time 20% drop in traffic 100 ms latency on page load 7% lower conversion rate

Slide 9

Slide 9 text

#Quarkus @holly_cummins why optimise? 0.5s extra search page time 20% drop in traffic 10 ms delay in trading platform 100 ms latency on page load 7% lower conversion rate

Slide 10

Slide 10 text

#Quarkus @holly_cummins why optimise? 0.5s extra search page time 20% drop in traffic 10 ms delay in trading platform 10% drop in revenue 100 ms latency on page load 7% lower conversion rate

Slide 11

Slide 11 text

#Quarkus @holly_cummins what is optimising?

Slide 12

Slide 12 text

#Quarkus @holly_cummins for whom? when? doing what? “make it go faster”

Slide 13

Slide 13 text

#Quarkus @holly_cummins human-centric (technical) design

Slide 14

Slide 14 text

#Quarkus @holly_cummins

Slide 15

Slide 15 text

#Quarkus @holly_cummins performance can be:

Slide 16

Slide 16 text

#Quarkus @holly_cummins performance can be: throughput

Slide 17

Slide 17 text

#Quarkus @holly_cummins performance can be: throughput transactions per second

Slide 18

Slide 18 text

#Quarkus @holly_cummins performance can be: throughput latency transactions per second

Slide 19

Slide 19 text

#Quarkus @holly_cummins performance can be: throughput latency transactions per second start-up time

Slide 20

Slide 20 text

#Quarkus @holly_cummins performance can be: throughput latency transactions per second response time start-up time

Slide 21

Slide 21 text

#Quarkus @holly_cummins performance can be: throughput latency ramp-up time transactions per second response time start-up time

Slide 22

Slide 22 text

#Quarkus @holly_cummins performance can be: throughput latency capacity ramp-up time transactions per second response time start-up time

Slide 23

Slide 23 text

#Quarkus @holly_cummins performance can be: throughput latency capacity ramp-up time transactions per second response time start-up time bandwidth

Slide 24

Slide 24 text

#Quarkus @holly_cummins performance can be: throughput latency capacity ramp-up time transactions per second response time start-up time footprint bandwidth

Slide 25

Slide 25 text

#Quarkus @holly_cummins performance can be: throughput latency capacity ramp-up time transactions per second response time start-up time CPU usage footprint bandwidth

Slide 26

Slide 26 text

#Quarkus @holly_cummins performance can be: throughput latency capacity utilisation ramp-up time transactions per second response time start-up time CPU usage footprint bandwidth

Slide 27

Slide 27 text

#Quarkus @holly_cummins performance can be: throughput latency capacity utilisation … ramp-up time transactions per second response time start-up time CPU usage footprint bandwidth

Slide 28

Slide 28 text

#Quarkus @holly_cummins performance can be: throughput latency capacity utilisation … ramp-up time transactions per second response time start-up time CPU usage footprint bandwidth

Slide 29

Slide 29 text

#Quarkus @holly_cummins Never underestimate the bandwidth [throughput] of a station wagon full of tapes hurtling down the highway. –Andrew Tanenbaum, 1981

Slide 30

Slide 30 text

#Quarkus @holly_cummins Never underestimate the bandwidth [throughput] of a station wagon full of tapes hurtling down the highway. –Andrew Tanenbaum, 1981 but the latency is terrible …

Slide 31

Slide 31 text

@holly_cummins requirements change

Slide 32

Slide 32 text

@holly_cummins

Slide 33

Slide 33 text

@holly_cummins

Slide 34

Slide 34 text

@holly_cummins

Slide 35

Slide 35 text

@holly_cummins

Slide 36

Slide 36 text

@holly_cummins I am not designed for this.

Slide 37

Slide 37 text

@holly_cummins the world changes

Slide 38

Slide 38 text

#Quarkus @holly_cummins

Slide 39

Slide 39 text

#Quarkus @holly_cummins -Xmx == $

Slide 40

Slide 40 text

#Quarkus @holly_cummins -Xmx == $ footprint

Slide 41

Slide 41 text

#Quarkus @holly_cummins

Slide 42

Slide 42 text

#Quarkus @holly_cummins which performs better?

Slide 43

Slide 43 text

#Quarkus @holly_cummins quarkus trading-off flexibility against startup speed and footprint

Slide 44

Slide 44 text

#Quarkus @holly_cummins quarkus trading-off flexibility against startup speed and footprint uhh … are you supposed to shut down applications after using them?

Slide 45

Slide 45 text

#Quarkus @holly_cummins which is better?

Slide 46

Slide 46 text

#Quarkus @holly_cummins which is better? which is faster?

Slide 47

Slide 47 text

#Quarkus @holly_cummins which is better? GraalVM Quarkus Application which is faster?

Slide 48

Slide 48 text

#Quarkus @holly_cummins which is better? OpenJDK GraalVM Quarkus Quarkus Application Application which is faster?

Slide 49

Slide 49 text

#Quarkus @holly_cummins which is better? ephemeral or serverless OpenJDK GraalVM Quarkus Quarkus Application Application which is faster?

Slide 50

Slide 50 text

#Quarkus @holly_cummins which is better? ephemeral or serverless OpenJDK GraalVM Quarkus Quarkus Application Application running your application for a long time which is faster?

Slide 51

Slide 51 text

#Quarkus @holly_cummins 30% of VMs are zombies (antithesisgroup.com) behaviour at idle

Slide 52

Slide 52 text

#Quarkus @holly_cummins how to optimise?

Slide 53

Slide 53 text

#Quarkus @holly_cummins find the bottleneck. fix it.

Slide 54

Slide 54 text

#IBM @holly_cummins what could possibly go wrong?!

Slide 55

Slide 55 text

#IBM @holly_cummins intuition pitfall 1

Slide 56

Slide 56 text

#Quarkus @holly_cummins this is not the place for ideas

Slide 57

Slide 57 text

#Quarkus @holly_cummins measure, don’t guess.

Slide 58

Slide 58 text

#Quarkus @holly_cummins measure the right thing

Slide 59

Slide 59 text

#Quarkus @holly_cummins measure the right thing what do your users care about?

Slide 60

Slide 60 text

#IBM @holly_cummins numbers pitfall 2

Slide 61

Slide 61 text

#Quarkus @holly_cummins

Slide 62

Slide 62 text

#Quarkus @holly_cummins leading indicators

Slide 63

Slide 63 text

#Quarkus @holly_cummins leading indicators lagging indicators

Slide 64

Slide 64 text

#Quarkus @holly_cummins leading indicators we care about them lagging indicators

Slide 65

Slide 65 text

#Quarkus @holly_cummins leading indicators we care about them easy to measure lagging indicators

Slide 66

Slide 66 text

#Quarkus @holly_cummins leading indicators we care about them easy to measure hard to change lagging indicators

Slide 67

Slide 67 text

#Quarkus @holly_cummins leading indicators we care about them easy to measure hard to change lagging indicators easy to change

Slide 68

Slide 68 text

#Quarkus @holly_cummins leading indicators we care about them easy to measure hard to change lagging indicators predictive of a thing we care about easy to change

Slide 69

Slide 69 text

#Quarkus @holly_cummins leading indicators we care about them easy to measure hard to change lagging indicators predictive of a thing we care about hard to identify easy to change

Slide 70

Slide 70 text

#Quarkus @holly_cummins leading indicators we care about them easy to measure hard to change lagging indicators predictive of a thing we care about hard to identify easy to change

Slide 71

Slide 71 text

#Quarkus @holly_cummins caution: performance experiments for entertainment purposes only. do not try these at home.

Slide 72

Slide 72 text

#Quarkus @holly_cummins 2007

Slide 73

Slide 73 text

#Quarkus @holly_cummins bad-ish advice: “reduce time spent in garbage collection”

Slide 74

Slide 74 text

#Quarkus @holly_cummins bad-ish advice: “reduce time spent in garbage collection” actually, garbage collection can make your application go faster

Slide 75

Slide 75 text

#Quarkus @holly_cummins 2007

Slide 76

Slide 76 text

#Quarkus @holly_cummins 2007

Slide 77

Slide 77 text

#Quarkus @holly_cummins 2021

Slide 78

Slide 78 text

#Quarkus @holly_cummins 2021

Slide 79

Slide 79 text

#Quarkus @holly_cummins -verbose:gc -Xverbosegclog:gclog.xml -Xcompactgc

Slide 80

Slide 80 text

#Quarkus @holly_cummins -verbose:gc -Xverbosegclog:gclog.xml -Xgcpolicy:optthruput -Xcompactgc

Slide 81

Slide 81 text

#Quarkus @holly_cummins -verbose:gc -Xverbosegclog:gclog.xml -Xgcpolicy:optthruput -Xmx110m -Xms110m -Xnocompactgc

Slide 82

Slide 82 text

#Quarkus @holly_cummins -verbose:gc -Xverbosegclog:gclog.xml -Xgcpolicy:optthruput -Xmx160m -Xms160m -Xnocompactgc

Slide 83

Slide 83 text

#Quarkus @holly_cummins -verbose:gc -Xverbosegclog:gclog.xml -Xgcpolicy:optthruput -Xmx300m -Xms300m -Xcompactgc why does the performance stay exactly the same no matter what gc settings I choose?

Slide 84

Slide 84 text

#Quarkus @holly_cummins by the way, this is cheating. (remember the ‘bad science’?)

Slide 85

Slide 85 text

#Quarkus @holly_cummins -verbose:gc -Xverbosegclog:gclog.xml -Xgcpolicy:optthruput

Slide 86

Slide 86 text

#Quarkus @holly_cummins -verbose:gc -Xverbosegclog:gclog.xml -Xgcpolicy:optthruput tool: GCMV

Slide 87

Slide 87 text

#Quarkus @holly_cummins -verbose:gc -Xverbosegclog:gclog.xml -Xgcpolicy:optthruput total GC time: 21.6s tool: GCMV

Slide 88

Slide 88 text

#Quarkus @holly_cummins -verbose:gc -Xverbosegclog:gclog.xml -Xgcpolicy:optthruput total GC time: 21.6s 4.1% of time in GC pause tool: GCMV

Slide 89

Slide 89 text

#Quarkus @holly_cummins -verbose:gc -Xverbosegclog:gclog.xml -Xgcpolicy:optthruput total GC time: 21.6s 4.1% of time in GC pause total GC time: 12.0s tool: GCMV

Slide 90

Slide 90 text

#Quarkus @holly_cummins -verbose:gc -Xverbosegclog:gclog.xml -Xgcpolicy:optthruput total GC time: 21.6s 4.1% of time in GC pause total GC time: 12.0s 3.6% of time in GC pause tool: GCMV

Slide 91

Slide 91 text

#Quarkus @holly_cummins -verbose:gc -Xverbosegclog:gclog.xml -Xgcpolicy:optthruput total GC time: 21.6s 4.1% of time in GC pause 23.9 GB garbage collected total GC time: 12.0s 3.6% of time in GC pause 13.0 GB garbage collected tool: GCMV

Slide 92

Slide 92 text

#Quarkus @holly_cummins -verbose:gc -Xverbosegclog:gclog.xml -Xgcpolicy:optthruput total GC time: 21.6s 4.1% of time in GC pause 23.9 GB garbage collected 493 transactions/s total GC time: 12.0s 3.6% of time in GC pause 13.0 GB garbage collected 260 transactions/s tool: GCMV

Slide 93

Slide 93 text

Slide 94

Slide 94 text

Slide 95

Slide 95 text

#Quarkus @holly_cummins total GC time: 21.6s 4.1% of time in GC pause 23.9 GB garbage collected 493 transactions/s total GC time: 12.0s 3.6% of time in GC pause 13.0 GB garbage collected 260 transactions/s

Slide 96

Slide 96 text

Slide 97

Slide 97 text

Slide 98

Slide 98 text

Slide 99

Slide 99 text

Slide 100

Slide 100 text

Slide 101

Slide 101 text

Slide 102

Slide 102 text

#Quarkus @holly_cummins so wait, what changed to make the app faster? running jmeter on the same machine as the app gives a big speedup!

Slide 103

Slide 103 text

#Quarkus @holly_cummins “Any improvements made anywhere besides the bottleneck are an illusion.” – Gene Kim

Slide 104

Slide 104 text

#Quarkus @holly_cummins time kills all performance advice (even mine)

Slide 105

Slide 105 text

#Quarkus @holly_cummins the takeaways: gc can improve performance by rearranging the heap find the bottleneck validate advice independently

Slide 106

Slide 106 text

#IBM @holly_cummins advice pitfall 3

Slide 107

Slide 107 text

#Quarkus @holly_cummins I read it on the internet!

Slide 108

Slide 108 text

#IBM @holly_cummins noooooo! “make one big method because method dispatching is slow”

Slide 109

Slide 109 text

#IBM @holly_cummins noooooo! “re-use your objects to help the garbage collector”

Slide 110

Slide 110 text

#IBM @holly_cummins noooooo! “to tune your JVM, use this command-line:” -server -Xms1g -Xmx1g -XX:PermSize=1g -XX:MaxPermSize=256m -Xmn256m -Xss64k -XX:SurvivorRatio=30 -XX:+UseConcMarkSweepGC -XX: +CMSParallelRemarkEnabled -XX:+UseCMSInitiatingOccupancyOnly -XX:CMSInitiatingOccupancyFraction=10 -XX:+ScavengeBeforeFullGC -XX: +CMSScavengeBeforeRemark -XX:+PrintGCDateStamps -verbose:gc -XX: +PrintGCDetails -Dsun.net.inetaddr.ttl=5 -XX:+HeapDumpOnOutOfMemoryError -XX:HeapDumpPath=`date`.hprof -Dcom.sun.management.jmxremote.port=5616 -Dcom.sun.management.jmxremote.authenticate=false -Dcom.sun.management.jmxremote.ssl=false -server -Xms2g -Xmx2g -XX:MaxPermSize=256m -XX:NewRatio=1 -XX:+UseConcMarkSweepGC

Slide 111

Slide 111 text

#IBM @holly_cummins noooooo! use StringBuilder, never concatenate strings with +=

Slide 112

Slide 112 text

#IBM @holly_cummins noooooo! use StringBuilder, never concatenate strings with += wait, what? yes, right?

Slide 113

Slide 113 text

#Quarkus @holly_cummins 2 things ruin advice: • context • time

Slide 114

Slide 114 text

#IBM @holly_cummins micro-optimisation pitfall 4

Slide 115

Slide 115 text

#Quarkus @holly_cummins

Slide 116

Slide 116 text

#Quarkus @holly_cummins static string beSlow() { string result = ""; for (int i = 0; i < 314159; i++) { result += getStringData(i); } return result; }

Slide 117

Slide 117 text

#Quarkus @holly_cummins @Override public String toString() { String ret = "\n\tMarket Summary at: " + getSummaryDate() + “\n\t\t TSIA:" + getTSIA() + “\n\t\t openTSIA:" + getOpenTSIA() + “\n\t\t gain:" + getGainPercent() + “\n\t\t volume:" + getVolume(); if ((getTopGainers() == null) || (getTopLosers() == null)) { return ret; } ret += "\n\t\t Current Top Gainers:"; Iterator it = getTopGainers().iterator(); while (it.hasNext()) { QuoteDataBean quoteData = it.next(); ret += ("\n\t\t\t" + quoteData.toString()); } ret += "\n\t\t Current Top Losers:"; it = getTopLosers().iterator(); while (it.hasNext()) { QuoteDataBean quoteData = it.next(); ret += ("\n\t\t\t" + quoteData.toString()); } return ret; }

Slide 118

Slide 118 text

Slide 119

Slide 119 text

Slide 120

Slide 120 text

Slide 121

Slide 121 text

#Quarkus @holly_cummins let’s make travel energy-efficient?

Slide 122

Slide 122 text

#Quarkus @holly_cummins every little helps?

Slide 123

Slide 123 text

#Quarkus @holly_cummins every little helps? every optimisation is another optimisation you aren’t doing

Slide 124

Slide 124 text

#Quarkus @holly_cummins our platforms help

Slide 125

Slide 125 text

#Quarkus @holly_cummins static string beSlow() { string result = ""; for (int i = 0; i < 314159; i++) { result += getStringData(i); } return result; }

Slide 126

Slide 126 text

#Quarkus @holly_cummins static string beSlow() { string result = “"; result += getStringData(1); result += getStringData(2); result += getStringData(3); return result; }

Slide 127

Slide 127 text

#Quarkus @holly_cummins static string beSlow() { string result = “"; result += getStringData(1); result += getStringData(2); result += getStringData(3); return result; } this is fine

Slide 128

Slide 128 text

#Quarkus @holly_cummins the JVM writers have far more time for optimising than you do clean, typical, code runs best

Slide 129

Slide 129 text

#Quarkus @holly_cummins ok, but how to optimise?

Slide 130

Slide 130 text

#Quarkus @holly_cummins tools

Slide 131

Slide 131 text

#Quarkus @holly_cummins “What you can optimize is limited to what you can observe.” -Susie Xia, Netflix

Slide 132

Slide 132 text

#Quarkus @holly_cummins observability

Slide 133

Slide 133 text

#Quarkus @holly_cummins method profiler GC analysis heap analysis APM distributed tracing * not free this is an incomplete list, because there are a lot of tools out there, and many cost money

Slide 134

Slide 134 text

#Quarkus @holly_cummins method profiler GC analysis heap analysis APM distributed tracing VisualVM * not free this is an incomplete list, because there are a lot of tools out there, and many cost money

Slide 135

Slide 135 text

#Quarkus @holly_cummins method profiler GC analysis heap analysis APM distributed tracing Mission Control VisualVM * not free this is an incomplete list, because there are a lot of tools out there, and many cost money

Slide 136

Slide 136 text

Slide 137

Slide 137 text

#Quarkus @holly_cummins method profiler GC analysis heap analysis APM distributed tracing Mission Control flame graphs VisualVM * not free this is an incomplete list, because there are a lot of tools out there, and many cost money IBM Health Center (for OpenJ9)

Slide 138

Slide 138 text

#Quarkus @holly_cummins method profiler GC analysis heap analysis APM distributed tracing Mission Control flame graphs GCMV VisualVM * not free this is an incomplete list, because there are a lot of tools out there, and many cost money IBM Health Center (for OpenJ9)

Slide 139

Slide 139 text

#Quarkus @holly_cummins method profiler GC analysis heap analysis APM distributed tracing Mission Control flame graphs GCMV VisualVM Eclipse MAT * not free this is an incomplete list, because there are a lot of tools out there, and many cost money IBM Health Center (for OpenJ9)

Slide 140

Slide 140 text

Slide 141

Slide 141 text

#Quarkus @holly_cummins method profiler GC analysis heap analysis APM distributed tracing Mission Control flame graphs GCMV New Relic* VisualVM Eclipse MAT * not free this is an incomplete list, because there are a lot of tools out there, and many cost money GlowRoot IBM Health Center (for OpenJ9)

Slide 142

Slide 142 text

#Quarkus @holly_cummins method profiler GC analysis heap analysis APM distributed tracing Mission Control flame graphs GCMV New Relic* AppDynamics* VisualVM Eclipse MAT * not free this is an incomplete list, because there are a lot of tools out there, and many cost money GlowRoot IBM Health Center (for OpenJ9)

Slide 143

Slide 143 text

#Quarkus @holly_cummins method profiler GC analysis heap analysis APM distributed tracing Mission Control flame graphs GCMV New Relic* AppDynamics* VisualVM Dynatrace* Eclipse MAT * not free this is an incomplete list, because there are a lot of tools out there, and many cost money GlowRoot IBM Health Center (for OpenJ9)

Slide 144

Slide 144 text

#Quarkus @holly_cummins method profiler GC analysis heap analysis APM distributed tracing Mission Control flame graphs GCMV New Relic* AppDynamics* Zipkin VisualVM Dynatrace* Eclipse MAT * not free this is an incomplete list, because there are a lot of tools out there, and many cost money GlowRoot IBM Health Center (for OpenJ9)

Slide 145

Slide 145 text

#Quarkus @holly_cummins method profiler GC analysis heap analysis APM distributed tracing Mission Control flame graphs GCMV New Relic* AppDynamics* Jaeger Zipkin VisualVM Dynatrace* Eclipse MAT * not free this is an incomplete list, because there are a lot of tools out there, and many cost money GlowRoot IBM Health Center (for OpenJ9)

Slide 146

Slide 146 text

#Quarkus @holly_cummins Netflix microservice architecture optimising a micro-service: is that micro-optimising?

Slide 147

Slide 147 text

@holly_cummins you may need to know the whole system context to know what to optimise

Slide 148

Slide 148 text

#Quarkus @holly_cummins don’t forget the edges queueing theory helps us understand where the disasters happen

Slide 149

Slide 149 text

#Quarkus @holly_cummins “When it comes to IT performance, amateurs look at averages. Professionals look at distributions.” – Avishai Ish-Shalom

Slide 150

Slide 150 text

#Quarkus @holly_cummins “Nines don’t matter if your users aren’t happy.” – Charity Majors

Slide 151

Slide 151 text

#Quarkus @holly_cummins slow performance can turn into big cloud bills make cloud costs visible to engineers

Slide 152

Slide 152 text

#Quarkus @holly_cummins ok, but you promised bears

Slide 153

Slide 153 text

#Quarkus @holly_cummins if you leave the TV on when you’re not using it, you’re a polar bear murderer

Slide 154

Slide 154 text

#Quarkus @holly_cummins there is a moral imperative to avoid waste

Slide 155

Slide 155 text

#Quarkus @holly_cummins there is a moral imperative to avoid waste electricity hardware

Slide 156

Slide 156 text

#Quarkus @holly_cummins data centres use 1-2% of the world’s electricity

Slide 157

Slide 157 text

#Quarkus @holly_cummins fewer devices longer lifetime

Slide 158

Slide 158 text

#Quarkus @holly_cummins fewer devices longer lifetime higher efficiency

Slide 159

Slide 159 text

#Quarkus @holly_cummins fewer devices longer lifetime higher efficiency lower footprint

Slide 160

Slide 160 text

#Quarkus @holly_cummins fewer devices longer lifetime higher efficiency lower footprint more multitenancy

Slide 161

Slide 161 text

#Quarkus @holly_cummins fewer devices longer lifetime higher efficiency lower footprint more multitenancy optimise for longevity

Slide 162

Slide 162 text

#Quarkus @holly_cummins fewer devices longer lifetime higher efficiency lower footprint more multitenancy optimise for longevity the end of planned obsolescence?

Slide 163

Slide 163 text

#Quarkus @holly_cummins sooo … you can optimise, and it can be fun measure, don’t guess only optimise what matters now for questions! @holly_cummins #Quarkus