How We Benchmarked Quarkus: Patterns and anti-patterns

Embed

Start on current slide

Slide 1

Slide 1 text

Holly Cummins Eric Deandrea Francesco Nigro How We Benchmarked Quarkus Patterns and anti-patterns Quarkus Insights April 20, 2026

Slide 2

Slide 2 text

@holly_cummins @edeandrea tl;dr

Slide 3

Slide 3 text

Quarkus is fast + efficient.

Slide 4

Slide 4 text

No content

Slide 5

Slide 5 text

quarkusio/spring-quarkus-perf-comparison

Slide 6

Slide 6 text

@holly_cummins @edeandrea The backstory

Slide 7

Slide 7 text

@holly_cummins @edeandrea The backstory What problem were we solving?

Slide 8

Slide 8 text

before

Slide 9

Slide 9 text

before what Quarkus version is this?

Slide 10

Slide 10 text

before what Quarkus version is this? when was it measured?

Slide 11

Slide 11 text

before what Quarkus version is this? where is the source code? when was it measured?

Slide 12

Slide 12 text

before what Quarkus version is this? where is the source code? when was it measured? how can I reproduce?

Slide 13

Slide 13 text

before what Quarkus version is this? where is the source code? when was it measured? how can I reproduce? what framework is this?

Slide 14

Slide 14 text

before

Slide 15

Slide 15 text

before why is throughput missing?

Slide 16

Slide 16 text

before why is throughput missing? is throughput so bad Quarkus had to hide the numbers?

Slide 17

Slide 17 text

No content

Slide 18

Slide 18 text

shadow-benchmarking

Slide 19

Slide 19 text

No content

Slide 20

Slide 20 text

home-rolled benchmark home-rolled benchmark home-rolled benchmark home-rolled benchmark home-r benchm home-rolled benchmark

Slide 21

Slide 21 text

duplicated effort benchmarking anti-patterns

Slide 22

Slide 22 text

We needed a benchmark.

Slide 23

Slide 23 text

We made a benchmark.

Slide 24

Slide 24 text

after

Slide 25

Slide 25 text

open source after

Slide 26

Slide 26 text

traceable open source after

Slide 27

Slide 27 text

@holly_cummins @edeandrea The backstory What could possibly go wrong?

Slide 28

Slide 28 text

No content

Slide 29

Slide 29 text

Building a benchmark is easy.

Slide 30

Slide 30 text

Building a benchmark is easy. Building a good benchmark is hard.

Slide 31

Slide 31 text

No content

Slide 32

Slide 32 text

No content

Slide 33

Slide 33 text

No content

Slide 34

Slide 34 text

No content

Slide 35

Slide 35 text

Benchmarks are like puppies

Slide 36

Slide 36 text

Benchmarks are like puppies

Slide 37

Slide 37 text

No content

Slide 38

Slide 38 text

No content

Slide 39

Slide 39 text

No content

Slide 40

Slide 40 text

No content

Slide 41

Slide 41 text

decisions,

Slide 42

Slide 42 text

decisions, decisions,

Slide 43

Slide 43 text

decisions, decisions, decisions

Slide 44

Slide 44 text

No content

Slide 45

Slide 45 text

decisions

Slide 46

Slide 46 text

decisions application code

Slide 47

Slide 47 text

decisions application code

Slide 48

Slide 48 text

decisions application code how the app is executed

Slide 49

Slide 49 text

decisions application code how the app is executed execution environment

Slide 50

Slide 50 text

decisions application code how the app is executed execution environment

Slide 51

Slide 51 text

Every decision changes the numbers.

Slide 52

Slide 52 text

Every decision changes the numbers.

Slide 53

Slide 53 text

How should we make decisions?

Slide 54

Slide 54 text

How should we make decisions?

Slide 55

Slide 55 text

Guiding Principles

Slide 56

Slide 56 text

Guiding Principles Parity Like-for-like comparison App code should be equivalent

Slide 57

Slide 57 text

Guiding Principles Parity Like-for-like comparison App code should be equivalent Normalness Representative of a typical app No occult tuning

Slide 58

Slide 58 text

Guiding Principles Parity Like-for-like comparison App code should be equivalent Normalness Representative of a typical app No occult tuning High quality Model best app dev practices Model best performance practices

Slide 59

Slide 59 text

Slide 60

Slide 60 text

reproducibility - can we get the same answer repeatedly? - is there noise in the results from things outside our control? - are we reporting useful metrics? - does this help us make a decision? - is it answering a question we actually care about? - is this close to real-world? - is this representative of the way applications will be run? realism relevance

Slide 61

Slide 61 text

realism reproducibility relevance - best practices for performance testing - parity (ie fairness) - test the framework, not the infrastructure - normalness - best practices for application code

Slide 62

Slide 62 text

It’s easy to make all three worse realism reproducibility relevance

Slide 63

Slide 63 text

For improvements, choose one :( reproducibility realism relevance

Slide 64

Slide 64 text

No content

Slide 65

Slide 65 text

echo "1" > /sys/devices/system/cpu/intel_pstate/no_turbo … —cpuset-cpus 1,4,6,8

Slide 66

Slide 66 text

echo "1" > /sys/devices/system/cpu/intel_pstate/no_turbo … —cpuset-cpus 1,4,6,8 But Francesco, no one would run a real application like that?

Slide 67

Slide 67 text

echo "1" > /sys/devices/system/cpu/intel_pstate/no_turbo … —cpuset-cpus 1,4,6,8 But Francesco, no one would run a real application like that? reproducibility realism relevance

Slide 68

Slide 68 text

Benchmarking is a post-truth discipline. realism reproducibility relevance

Slide 69

Slide 69 text

@holly_cummins @edeandrea Some of our learnings

Slide 70

Slide 70 text

@holly_cummins @edeandrea Mistakes we’ve made

Slide 71

Slide 71 text

@holly_cummins @edeandrea Decisions we’ve made

Slide 72

Slide 72 text

reproducibility realism relevance out-of-the-box or tuned?

Slide 73

Slide 73 text

reproducibility realism relevance how long should warmup be?

Slide 74

Slide 74 text

reproducibility realism relevance pinning work to cores

Slide 75

Slide 75 text

transactional reads? reproducibility realism relevance

Slide 76

Slide 76 text

reproducibility realism relevance database container networking

Slide 77

Slide 77 text

reproducibility realism relevance deterministic cpu frequency (spoiler: it isn’t by default)

Slide 78

Slide 78 text

reproducibility realism relevance how much data should be in the database?

Slide 79

Slide 79 text

reproducibility realism relevance stale database images (oops)

Slide 80

Slide 80 text

Decisions are hard (and unfair). Can we just … measure both ways?

Slide 81

Slide 81 text

No content

Slide 82

Slide 82 text

JVM Native

Slide 83

Slide 83 text

Spring 4 Spring 3

Slide 84

Slide 84 text

Out of the box Tuned

Slide 85

Slide 85 text

With Virtual Threads No Virtual Threads

Slide 86

Slide 86 text

AOT Normal

Slide 87

Slide 87 text

AOT Normal Leyden AOT

Slide 88

Slide 88 text

AOT Normal Leyden AOT Spring AOT

Slide 89

Slide 89 text

Combinatorics.

Slide 90

Slide 90 text

Our current measurement matrix is … a lot. Out-of-the-box Tuned

Slide 91

Slide 91 text

@holly_cummins @edeandrea How universal are the results?

Slide 92

Slide 92 text

“It depends.”

Slide 93

Slide 93 text

The ultimate validation

Slide 94

Slide 94 text

The ultimate validation They could run 3 times denser deployments without sacrificing availability and response times of services”, Thornsten reiterated.

Slide 95

Slide 95 text

No content

Slide 96

Slide 96 text

I tried your benchmark, and Quarkus is only 1.4x faster

Slide 97

Slide 97 text

decisions application code how the app is executed execution environment

Slide 98

Slide 98 text

decisions application code how the app is executed execution environment

Slide 99

Slide 99 text

decisions application code how the app is executed execution environment

Slide 100

Slide 100 text

decisions application code how the app is executed execution environment

Slide 101

Slide 101 text

hardware schedulers make a big difference

Slide 102

Slide 102 text

@holly_cummins @edeandrea The mistakes we didn’t make

Slide 103

Slide 103 text

@holly_cummins @edeandrea How not to benchmark

Slide 104

Slide 104 text

reproducibility realism relevance Running on a laptop don’t do this

Slide 105

Slide 105 text

reproducibility realism relevance don’t do this Other work running on the same machine (the load driver counts!)

Slide 106

Slide 106 text

reproducibility realism relevance Not having a clear question don’t do this

Slide 107

Slide 107 text

reproducibility realism relevance Not measuring what you think you’re measuring don’t do this

Slide 108

Slide 108 text

reproducibility realism relevance A measurement of the wrong bottleneck is a useless measurement

Slide 109

Slide 109 text

solution: active benchmarking

Slide 110

Slide 110 text

reproducibility realism relevance Coordinated omission don’t do this

Slide 111

Slide 111 text

reproducibility realism relevance Varying multiple things at once don’t do this

Slide 112

Slide 112 text

reproducibility realism relevance Measuring multiple things at once don’t do this