Concurrent programmin g - · PDF fileConcurrent programmin The industrial expe gxP rience. ... assignment at Ericsson ystem ge 2. ... alarms (e.g. heartbea

Concurrentprogrammin

The industrial expe

t ngxP

erience

Karol Ostrovský

› M.Sc. – Comenius University

› Ph.D. – Chalmers

› Post-doc – Chalmers

› System Designer – Dfind, on – Operations & Maintenance Subs

Guest Lecture PPxT VT2011 | Karol Ostrovsky | © Ericsson AB 2011 | 2011-02-21 | Pa

, Bratislava

assignment at Ericssonsystem

age 2

Karol Ostrovský

› Language skillsB i–Basic

–Pascal–C/C++–C/C++–Scheme–SmallTalkSmallTalk–Java–JR (MPD)–Haskell–Erlang


–OcamlLaTeX–LaTeX

–VAX assembler–TrilogyTrilogy–Ada–Agdag

age 3

The chalmers yea

› Research in static analysis oflanguageslanguages

– Type systems– Protocol analysis

› Main course responsible PPxDeveloped the course between– Developed the course between


ars

f concurrent programming

xT2005 and 20102005 and 2010

age 4

The case of the m

The Message from

• Should you forget evercourse, please, rememp

Use the right to


issing slide

m this Course

rything from this mber at least this saying:y g

ool for the job.

age 5

3PPVT10 – Introduction

What does this bu

Course Goals

• Introduction to the promany computing discipy p g p◦ Operating systems◦ Distributed systemsR l ti t◦ Real‐time systems

• Appreciation of the proprogrammingp g g◦ Classic synchronisation


uy you?

– General

oblems common to plines:p

oblems of concurrent

n problems

age 6


What does this bu

Course Goals

• Understanding of a range constructs for concurrent

• Ability to apply these in pproblems in concurrent pp p

• Practical knowledge of theof modern concurrent proof modern concurrent pro


uy you?

– Practical

of programming language programming

ractice to synchronisation rogrammingg g

e programming techniques ogramming languagesogramming languages

age 7


what does this re

› Everything this course teache– As is– As is

› Erlang lectures – almost prod

– Applied or customised› Resource allocation› Master – slave› Timeouts, alarms


eally buy you?

es is used in industrial practice

duction code

age 8

The pthreads expe

› SGSN shell in CProblem: Unix signal handling› Problem: Unix signal handling

› Solution: create a separate th– My problemy p

› I never practically used pthre– My solution

› general solution based on co› general solution based on co- synchronisation, mutual e

› “translate” to pthreads


erience

g was insufficientg was insufficienthread for signal handling

eads before

oncurrency principlesoncurrency principlesexclusion, message passing

age 9

The general solut

Main

Shared

Main

Handler


tion

d data

age 10

Pthreads transla

init

Main

Mutex vlock

Maincreate

unlockHandler


ation

destroyvariable unlock

join

age 11

Ericsson – where a

› Business Unit Networks– Development Unit IP and Broad– Development Unit IP and Broad

› Product Development Unit P- SGSN-MME

› O&M sub-system› 2G sub-system› 3G sub-systemy› ...

- GGSN-MPG- CPG- CPG- ...


am i?

bandbandacket Core

age 12

Mobile telecom n


etwork

age 13

packet core netw

› 3GPP– Defines standards (mostly proto– Defines standards (mostly proto– Interoperability is essential

› SGSN-MME– Servicing GPRS Support Node (– Mobility Management Entity (4GMobility Management Entity (4G– Control signalling

› Admission control, AuthenticM bilit i› Mobility, roaming

– Payload transport (not in 4G)


work

ocols)ocols)

(2G/3G)G)G)

ation

age 14

SGSN-MME – mkvi

› 3 sub-racks› 21 blades (2+19)› 2 core PowerPC› ~ 114 simultaneously running processes

› Backplane: 1Gbps

› Up to 3 MSAU

Guest Lecture PPxT VT2011 | Karol Ostrovsky | © Ericsson AB 2011 | 2011-02-21 | Paage 15

SGSN-MME – MkViii

› 3 sub-racks› 14 blades (2+12)› 6 core Intel x86

–12 SMT threads total› ~ 432 simultaneously

irunning processes

B k l 1 10Gb› Backplane: 1 or 10Gbps

Guest Lecture PPxT VT2011 | Karol Ostrovsky | © Ericsson AB 2011 | 2011-02-21 | Paage 16

SGSN-MME – archit

NCB FSNCBNCB FSNCB

AP APAP AP

DP DP


tecture sketch

SB FSBSB FSB

...... APAP

... DP

age 17

Sieve of Eratosth

Architec

• Onenums One

filter1 filter2filter1 filter2

outpprint


enes

cture

• N+1 pipeline channelse shared output channele shared output channel

filterNfilterN

eatput

age 18

46PPVT10 – Message Passing

pipeline of proces

NCNC

AP_1 AP_2

logLOG


g

sses

CBCB

AP_N

gging

age 19

gg g

Typical concurre

› One mobile – one process (re– Isolation– Isolation– Synchronisation only with resou

› Central resources– Resource allocator– Master/coordinator – slave/work

› Transaction handler› Transaction handler› Node management

– Configuration management– Fault management (e.g. barrier – Timeouts, alarms (e.g. heartbea


ncy patterns

eplicated worker)

rces

ker

sync. during start-up)at)

age 20

SGSN-MME

› A huge version of the assignm– The pub SGSN MME node– The pub – SGSN-MME node– Customers – mobile phones– The landlord – NCB– The assistant – node manageme– The tables – system resources

› Extension of the assignment– SGSN-MME can run in a pool –– Operations & Maintenance?


ment 2

ent

multiple-bars?

age 21

Operations & Main

› Configuration management– Maintaining working and consist– Maintaining working and consist

› Logging– “Black box” (flight data recorder)

› Fault management– Alarm handling

P f t› Performance management– Gathering statistics– Monitoring health


ntenance

tent settingstent settings

)

age 22

Haskell patterns

› Good– Keeps pure and side effecting c– Keeps pure and side-effecting c

› Good separation of concerns› Improved compositionality› Possible performance gain

– Gather writes together and writetransactions:

› 1 item write costs 10› 10 items write is not 100 but

› Bad› Bad– In rapid prototyping it can presen– So, it is good that Erlang does n


– monads

omputations apartomputations aparts

e to DB once – amortise the cost of

only 20!

nt a big hurdle to jump overnot have static types

age 23

OO-design pattern

› Factory method– Improve memory sharing– Improve memory sharing

› Object poolj p– Bounded parallelisation of algor– Overload protection


ns

ithms – thread pool

age 24

What they do not

› Software maintenance– Software lives long– Software lives long

› Especially telecom systems (› Banking systems live even lo

– People change– Organisations change– Hardware changesg– Requirements change– Documentation often does not c


t teach you

(decades)onger (think COBOL)

change

age 25

Software mainten

› The developer’s challenge– Write simple (readable) and effic– Write simple (readable) and effic

1. Write a straightforward and2. Optimise later (or even bet

› Think smart but do not over-oOptimisations complicate mainte– Optimisations complicate mainte

› The code is often the only rely– Types can be very good docume


nance

cient code:cient code:d working solution firsttter skip this step)

optimiseenanceenance

iable documententation

age 26

Synthesis and ana

› Emphasis on synthesis so far– Software development– Software development

› Perhaps 30% of manpower isp p– Analytical work– Do you like to break a system?

B t t ti l b “ th› But testing can also be “synth– Testing frameworks

› Quickcheck› SGSN-MME has its own

– Would you like to formally prove


alysis

r

s spent on testingp g

h ti l”hetical”

e the system correct?

age 27

Erlang in practic

› Well suited for– Control handling of telecom traff– Control handling of telecom traff– Application layer (OSI model) ap

› Web servers, etc.– Domain Specific Language – fra

› Test scripting

› Reasonably high-level (as co› Reasonably high level (as co– Good for software maintenance

› OTP – includes useful buildin– Supervisor– Generic server– Finite state machine


e – pros

ficficpplications

amework

mpared to for example C)mpared to for example C)

g blocks

age 28

Erlang in practic

› Hard to find good Erlang prog

› A bit too low-level language– Given current HW limitations on

point where the code is not porta– Raise the abstraction and provid

› Where is the type system?– A static type system of Haskell-n– But good static analysis tools ar


e – cons

grammers (?)

e must sometimes optimise to the pable (with the same performance)de a customisable compiler, VM

nature would be a hindrancere desperately needed

age 29

More than true

Saying

• The greatest performais when a system goes y gworking

• The only thing worse thappens all the time ishappen all the timepp


gs

ance improvement of all from not‐working to g

han a problem that s a problem that doesn't

age 30


Concurrent prog

› You are finishing a very good– Provides great general/theoretic– Provides great general/theoretic– Mixed with almost production lev

› The industrial experience– It is more difficult that you expec

Good systematic approach help– Good systematic approach help– Lateral thinking is an asset


ramming xp

coursecal backgroundcal backgroundvel practical examples

ct, butss

age 31

Concurrent programmin g - · PDF fileConcurrent programmin The industrial expe gxP rience. ... assignment at Ericsson ystem ge 2. ... alarms (e.g. heartbea

Documents