Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
Concurrency Basics for Elixir
Search
Maciej Kaszubowski
August 02, 2018
Programming
0
120
Concurrency Basics for Elixir
Slides from internal presentation at
https://appunite.com
Maciej Kaszubowski
August 02, 2018
Tweet
Share
More Decks by Maciej Kaszubowski
See All by Maciej Kaszubowski
Error-free Elixir
mkaszubowski
0
340
Modular Design in Elixir (ElixirConf EU 2019)
mkaszubowski
2
740
The Big Ball of Nouns
mkaszubowski
0
110
Modular Design in Elixir
mkaszubowski
1
390
Our three years with Elixir
mkaszubowski
0
250
Distributed Elixir
mkaszubowski
0
160
Software Architecture
mkaszubowski
0
140
Let it crash - fault tolerance in Elixir/OTP
mkaszubowski
0
480
CRDTs - The science behind Phoenix Presence
mkaszubowski
2
270
Other Decks in Programming
See All in Programming
Deep Dive into Kotlin Flow
jmatsu
1
360
スケールする組織の実現に向けた インナーソース育成術 - ISGT2025
teamlab
PRO
1
150
私の後悔をAWS DMSで解決した話
hiramax
4
210
Zendeskのチケットを Amazon Bedrockで 解析した
ryokosuge
3
310
Testing Trophyは叫ばない
toms74209200
0
890
ユーザーも開発者も悩ませない TV アプリ開発 ~Compose の内部実装から学ぶフォーカス制御~
taked137
0
190
より安全で効率的な Go コードへ: Protocol Buffers Opaque API の導入
shwatanap
2
630
GitHubとGitLabとAWS CodePipelineでCI/CDを組み比べてみた
satoshi256kbyte
4
250
Updates on MLS on Ruby (and maybe more)
sylph01
1
180
そのAPI、誰のため? Androidライブラリ設計における利用者目線の実践テクニック
mkeeda
2
1.8k
AI時代のUIはどこへ行く?
yusukebe
18
9k
AIと私たちの学習の変化を考える - Claude Codeの学習モードを例に
azukiazusa1
10
4.4k
Featured
See All Featured
ReactJS: Keep Simple. Everything can be a component!
pedronauck
667
120k
Visualizing Your Data: Incorporating Mongo into Loggly Infrastructure
mongodb
48
9.7k
Cheating the UX When There Is Nothing More to Optimize - PixelPioneers
stephaniewalter
285
14k
How to train your dragon (web standard)
notwaldorf
96
6.2k
Bash Introduction
62gerente
615
210k
The Language of Interfaces
destraynor
161
25k
Balancing Empowerment & Direction
lara
3
620
RailsConf & Balkan Ruby 2019: The Past, Present, and Future of Rails at GitHub
eileencodes
139
34k
Building Adaptive Systems
keathley
43
2.7k
Site-Speed That Sticks
csswizardry
10
820
Fashionably flexible responsive web design (full day workshop)
malarkey
407
66k
Principles of Awesome APIs and How to Build Them.
keavy
126
17k
Transcript
Concurrency basics For Elixir-based Systems
None
So, what’s concurrency?
Sequential Execution (3 functions, 1 thread)
Sequential Execution (3 functions, 1 thread) Concurrent Execution (3 functions,
3 threads)
Sequential Execution (3 functions, 1 thread) Concurrent Execution (3 functions,
3 threads) Preemptive scheduling
Where’s the benefit?
Req1 Req2 Req3 Resp Sequential Execution time Waiting time
Req1 Req2 Req3 Resp Req1 Resp Req2 Req3 Sequential Concurrent
Execution time Waiting time
CPU bound Re Re Re Res Re Res Re Re
I/O bound
Concurrent or Parallel What’s the difference?
Concurrent Execution (3 functions, 3 threads)
Concurrent Execution (3 functions, 3 threads) Parallel Execution (3 functions,
3 threads, 2 cores) core 1 core 2
root@kingschat-api-c8f8d6b76-4j65j:/app# nproc 12 root@tahmeel-api-prod-b5979bdc6-q5wz6:/# nproc 1 How many cores?
Concurrent Execution (3 functions, 3 threads) Parallel Execution (3 functions,
3 threads, 2 cores) core 1 core 2 (by default) One erlang scheduler per core
:observer_cli.start()
None
Req1 Req2 Req3 Resp Req1 Resp Req2 Req3 Sequential Concurrent
Execution time Waiting time Req1 Resp Req2 Req3 Parallel
Sequential execution
Phoenix Request Req 1
Phoenix Request Resp
Phoenix Request Req 2
Phoenix Request Resp
Phoenix Request Req 3
Phoenix Request Resp
Concurrent execution
Phoenix Request
Phoenix Request Task 1 Task 2 Task 3
Phoenix Request Task 1 Task 2 Task 3 Req 1
Req 2 Req 3
Phoenix Request Task 1 Task 2 Task 3 Resp Resp
Resp
Phoenix Request Task 1 Task 2 Task 3
R1 APP Server DB Server (3 cores) R2 R1 R2
Time Execution time Waiting time
R1 APP Server DB Server (3 cores) Send resp R2
R3 R1 R2 R3 Time Execution time Waiting time
How much can we gain?
Amdahl’s Law
Amdahl’s Law
Amdahl’s Law in a nutshell The more synchronisation, the less
benefit from multiple cores
R1 APP Server Send resp R2 R3 R1 R2 R3
Time Execution time Waiting time Almost 100% parallel (almost no synchronisation) DB Server (3 cores)
But…
R1 APP Server Send resp R2 R3 R1 R2 R3
Time Execution time Waiting time This is not constant DB Server (3 cores)
R1 APP Server Send resp R2 R3 R1 R2 R3
Time Execution time Waiting time This is not infinite DB Server (3 cores)
R1 APP Server R2 R3 R1 R2 R3 R4 Time
Execution time Waiting time DB Server (3 cores)
R1 APP Server R2 R3 R1 R2 R3 R4 Time
Execution time Waiting time DB Server (3 cores)
R1 APP Server R2 R3 R1 R2 R3 R4 R4
Time Execution time Waiting time DB Server (3 cores)
R1 APP Server R2 R3 R1 R2 R3 R4 R4
Time Execution time Waiting time DB Server (3 cores)
R1 APP Server R2 R3 R1 R2 R3 R4 R4
Time Execution time Waiting time DB Server (3 cores)
R1 APP Server Send resp R2 R3 R1 R2 R3
R4 R4 Time Execution time Waiting time DB Server (3 cores)
R1 APP Server R2 R3 R1 R2 R3 R4 R4
Time Execution time Waiting time R5 R6 R7 R5 R6 R7 DB Server (3 cores)
Phoenix Request Task 1 Task 2 Task 3 Req 1
Req 2 Req 3 Remember this?
This isn’t exactly true
None
Connection pool (Prevents from overworking the DB)
Pool Manager (Blocks until a free worker is available)
None
Pool Manager (Blocks until a free worker is available)
None
It gets worse
Pool Manager Mailbox Has to be synchronised
Pool Manager Message Passing Is just copying data in shared
memory
Pool Manager Remember semaphores?
Logger Metrics Sentry
Network stack
Network stack
Network stack
Network stack Sentry Metrics
OS Threads (Garbage Collection) Data Bus Virtual Machines Memory characteristics
(e.g. processor caches) … Other synchronisation points
That’s hard
That’s REALLY hard
That’s REALLY hard Seriously, people spend their entire careers on
this
So, what to do?
Measure
Measure Measure
Measure Measure Measure
Measure ON PRODUCTION
Measure ON PRODUCTION You WILL get false results on staging/locally
Measure Entire system You WILL get false results for single
functions
Measure ONLY IF YOU HAVE TRAFFIC
“premature optimization is the root of all evil”
If something takes X ms, it will always take X
ms.
Async execution cannot “remove” this time It can only hide
it
BACK PRESSURE
Producent Consumer Consumer
Producent Consumer Consumer
Producent Consumer Consumer
Producent Consumer Consumer
Producent Consumer Consumer
Producent Consumer Consumer
Producent Consumer Consumer
Producent Consumer Consumer
Producent Consumer Consumer
Producent Consumer Consumer
Producent Consumer Consumer
Producent Consumer Consumer
Producent Consumer Consumer
Producent Consumer Consumer Stop
Producent Consumer Consumer
Producent Consumer Consumer
Producent Consumer Consumer
Producent Consumer Consumer
Producent Consumer Consumer OK, give me more
Producent Consumer Consumer
None
Back pressure
Thanks!