Transcript 08-01-30_Kedar - Columbia University

A Whirlwind Tour through
Concurrency
Kedar Namjoshi
Bell Labs
Columbia University, Jan. 2008
1
Outline
 Concurrency
 Common Concurrency Patterns
 The Plan9 Thread Model
 Reasoning about Concurrency
 The “Session Data Type” Model
2
Concurrency
Processes which overlap in time and exchange information
 The world is concurrent. Purely sequential
behavior is often artificial (“like a machine”)
 However, writing concurrent programs is
viewed as a hard problem. Why?
3
Concurrency in programming
 For a long time, concurrency has been a concern
mainly for designers and implementers of
–
–
–
–
–
operating systems
supercomputers
processor chips
database systems
and phone networks …
This is about to change!
Faster Chips Are Leaving Programmers In Their Dust
(The New York Times, December 17, 2007)
4
Multi-Core Processors
(from http://www.amd.com)
Dual-core already common; Quad-core on its way; expect more!
5
Shared-Memory Concurrency
initially, account x holds $0
P1:
x := x+50; // deposit $50 to x
P2:
x := x+100; // deposit $100 to x
What is the amount in x after P1||P2?
It should be $150 --- but it might be $50 !
How could this happen?
6
Shared-Memory II
Let’s take a closer look
P1: t1 := x; t1 := t1+50; x := t1;
P2: t2 := x; t2 := t2+100; x := t2;
For the schedule P2;P1;P2;P1;P2;P1, the final value
of x is $50!
Mistake: implicit assumption that the deposit action is
atomic.
7
Shared-Memory III
Atomicity must be enforced.
P1’: lock x; deposit $50; unlock x
P2’: lock x; deposit $100; unlock x
Locks are difficult to get right.
 programmers forget to place them (“data races”)
 or forget to release locks, or do so in the wrong order (“deadlock”)
 or put too many locks (minimal concurrency)
 Locking could result in priority inversion (a higher priority process gets
locked out) --- the Mars Pathfinder failure was traced back to priority
inversion
Locks often don’t match up well to the structure of a problem.
8
Shared Memory IV
Compilers can get in the way, too!
P1:
P2:
x := 0;
y := 0;
y := 1;
x := 1;
On termination, we may expect anything but x=0,y=0. However, the
compiler (or the processor) is free to reorder instructions!
P1’:
y := 1;
x := 0;
P2’:
x := 1;
y := 0;
Now x=0,y=0 is a possibility!
This leads to considerations of a weak consistency model.
9
CSP [Hoare 1978]
CSP: Communicating Sequential Processes
Three key concepts
 No shared variables
 Processes communicate through synchronized send-receive
actions
 A process may respond to one of a set of communication
actions
Simple. Clean. Elegant.
Hoare received the Turing Award in 1980 for
this and many other contributions to computing
10
Communication Patterns
Often, the most natural description of a concurrent
program is through its communication pattern.
We’ll see a number of very common patterns.
11
1. The knock on the door …
… also known as a “real-time interrupt”
C + Unix : signals
Java: Thread.interrupt()
“As captain of the crew I had had extensive experience (dating back to 1958) in making
basic software dealing with real time interrupts and I knew by bitter experience that as a
result of the irreproducibility of the interrupt moments, a program error could present
itself misleadingly like an occasional machine malfunctioning. As a result I was terribly
afraid. Having fears regarding the possibility of debugging we decided to be as careful
as possible and —prevention is better than cure!— to try to prevent bugs from entering
the construction.”
-- Edsger W. Dijkstra, The Structure of the “THE”–multiprogramming System
(1968) EWD 196
Dijkstra received the Turing Award in 1972
for his work on concurrency and structured
programming
12
How does one prevent bugs?
The non-determinism introduced by the real-time
interrupt implies that a program execution may be
stopped at any point, in any intermediate state.
The key is to write down the minimal set of facts that
should be invariant (i.e., unchanged) across an
interrupt, for the main program to work correctly.
Then write the interrupt handling routine to respect
the invariant, or disable interrupts over a short
execution sequence.
13
2. Copy
process copy(channel in, channel out)
var x;
*[in?x  out!x]
? is the input action
! is the output action
*[] means repeat forever
14
3. Pipeline
A chain of N (N >= 1) copy processes:
copy(c0,c1) || copy(c1,c2) || … || copy(c(N-1),c(N))
If the i’th process is altered to
[in?x  out!f(i)(x)]*
the pipeline computes the function
f(N) * f(N-1) * … * f(2) * f(1)
15
4. Delegation: Taxi Rank
All taxis follow the same protocol.
process Taxi[i]:
taxi-rank! i; *[dispatch[i]?p  drive(p); taxi-rank! i]
process Dispatcher:
*[queue?p  (taxi-rank?i  dispatch[i]!p)]
Also known as a “thread pool”
16
5. Delegation: Futures
process A:
toB ! (“x+2”,x:10);
…. stuff ….
fromB ? y;
Process B:
*[toB? (expr, binding)  fromB! eval(expr,binding)]
Introduced in Scheme (1975?) as delay and force
17
6. Co-authoring a paper
process Kedar:
toDennis! token;
*[fromDennis? token  write; toDennis! token]
process Dennis:
*[toDennis? token  write; fromDennis! token]
Also known as a “co-routine” .
18
7. Callback
K:
C! “let me know if you can’t pick up the kids today”;
[work interrupt (C? “can’t do it”  pick up kids)]
Best understood as an expected interrupt.
All warnings regarding interrupt processing apply.
19
8. Multiplexing
Merge messages from two channels into a single channel
process mux:
*[
| channel1? x  out! x
| channel2? y  out! y
]
The operator “|” defines alternatives. If both alternatives are available,
the choice is made non-deterministically.
The result may not be a fair merge!
20
9. De-Multiplexing
Copy input from one channel into two
process demux:
*[ input? x  out1! x; out2! x]
Or, perhaps:
process demux:
*[input? X  if (x>0) out1!x; else out2!x]
21
10. On Friday …
K:
C! “let me know if you can’t pick up the kids today”;
[work interrupt
(
| C? “can’t do it”  pick up kids
| Al? “would you like to teach a class?”  chat
| Dennis? “lunch?”  go to lunch
|…
)
]
All warnings regarding interrupt processing apply.
22
The Plan9 Thread Model
Plan9 is an operating system developed at Bell Labs from about 1990 onwards.
Lots of interesting features.
In particular, the OS thread model influenced by CSP.
CSP also influenced two languages at Bell Labs
 newsqueak (Pike, Cardelli) Rob Pike’s talk
 Alef (Winterbottom) Russ Cox’s CSP-Plan9 page
and, in the wider world …
 The design of the Ada programming language
 The Occam programming language
 The Transputer chip
 The BPEL web services language
The CSP book is available.
23
What is the Plan9 thread model?
Two kinds of entities: processes and threads.
 Threads are co-routines within a process (i.e., scheduling is
cooperative)
 Processes are pre-emptively scheduled
 Channels are used for communication and synchronization.
They can be synchronous (as in CSP) or buffered (with finite
capacity, essentially a pipeline)
Further reading: these lecture notes (Sape Mullender) and a
description of the structure of the Plan9 window system (Pike)
24
How is it used?
 Typically, shared data is held in one process, and
accessed through commands sent through a
channel (e.g., read, write)
 Hence, no locks. Synchronization is enforced
through channels. Hence, no race conditions
 Clean fit to the design
Communication deadlock is possible, but it can often
be detected early by analyzing the communication
“skeleton”
25
Programming Languages: Sharedvariable concurrency
 High-Performance Fortran (HPF), OpenMP, X10,
C*: designed for high-performance, scientific
computing. “forall” construct, data layout
specification
 Java: Threads, synchronized methods, external
locks. Also, lock-free implementations of basic data
structures in java.util.concurrent
 C/C++: concurrency provided by external thread
libraries (e.g., pthreads). C++ will include language
constructs for concurrency in its next version
 OCaml/F#: OCaml and F# use external thread
libraries. F# has a concurrent garbage collector
26
Programming Languages: Messagepassing concurrency
 Ada: processes (called tasks) communicating
through “rendezvous” operations (like CSP ! and ?)
 Erlang: processes communicating though
unbounded message queues. Used by Ericsson for
programming telephony switches
 MPI: Message-passing interface. Used for widearea and supercomputer applications
27
Bell Labs Research
Session Data Types: Simple
Management of the Life-cycle
of Session-oriented Services
Al Aho*, Dennis Dams, Rick Hull, John Letourneau,
Kedar Namjoshi, Frans Panken, Mans van Tellingen
Alcatel-Lucent and *Columbia University
28
Session Data Types
A new project at Bell Labs
 a simple model and a (fairly simple) API for
programming real-time communications
 Targets applications that mix web-based setup and
control with real-time communications via telephony,
video, and IM
 Model based around the key concept of a session.
29
The “impedance mismatch”
Device OS’s
and
platforms
Devices
Eclipse-based SCE
SIP App
Server
(including
SB)
Web
Services
App
Server
Service
Developer/
Blender
SIP,
SIP
Servlets, SB
Steplets, …
Other?
...
Web Services,
SOAP, WSDL,
…
Programming of converged services: programmers still
think in multiple paradigms
– Web: HTTP, SOAP, WSDL, XML
– All-IP Telecom: SIP, IMS, Parlay/ParlayX, AIN, PacketTV,
XMPP, …
30
SDT Model
Three basic building blocks
Party: e.g., kedar@cellphone, dennis@IMserver, aho@laptop
Bubble: a group of parties in conversation. The group could be
empty, or have a single person in it (usually a temporary situation)
Session: a group of bubbles, related in some way
And three key concepts
Application: a control program
State Machine: Each party is limited to behavior specified by a
state machine over events such as “pickup”, “drop”, “mute”
Notify-Respond: the application is notified for events it
registers for. It can trigger changes to a state machine or to
a session in response
31
SDT Model
Session Manager
Session 1
Bubble 2
Bubble 1
Bob
Alice
Deborah
Charles
32
Global Address
Book /
Social Network
[ people,
devices,
friends ]
“Hello World” in SDT
// register as a client for a session manger
ISessionManager sm = SessionManager.register(“config.properties”,
this);
// start a session with a predefined audio bubble
AudioSDT s = new AudioSDT(sm);
AudioBubble b = s.getBubble();
// add yourself and a friend
b.addParty(new PartyID(“Kedar”, “19085821891”), Role.ReadWrite);
b.addParty(new PartyID(“Dennis”, “19085828859”), Role.ReadWrite);
// …… talk ……
System.out.println(“Hello World!”);
Thread.sleep(whatever);
// exit: sm shuts down sessions and bubbles
33
A bit more …
// set up a handler for pickup actions
public class pickupHandler implements IHandler {
public Response run(TriggerBinding[] b) {
num_pickup++;
return Response.Continue;
}
}
// link the handler to the pickup action
PartyVar who = new PartyVar(“who”);
Expr e = b.mkEvent(AudioBubble.SM.Events.pickup(),who);
b.addTrigger(e, new pickupHandler());
34
And that’s it!
The session/bubble model is simple, yet very flexible, thanks to
the notify-respond mechanism.
 Typically, the application program sets up sessions and
bubbles and waits for events to happen
 The response to an event may create (or delete) sessions
and bubbles, add (or drop) parties, or create new event
triggers, all at run-time
This results in an incredible amount of flexibility, out of a very
few basic primitives.
35
What we have done
We have an implementation of this model, running on
an externally available server
We have written sample applications, including:
 A customizable, network-hosted, speed-conference
Through a web interface, set your phone buttons to perform
SDT actions. As the service is network-hosted, these bindings
work from anywhere. [subject to some fine print]
 A conference system with the ability to create
“whisper” sub-sessions
36
What you might do
 Design a scripting language that hides some of the
Java intricacies
 Design a graphical, web-based language to
implement your own view of speed dial
 Create a conferencing application which you can
distribute through Facebook
 … or something far more interesting! 
We are available throughout the semester for help
with the API and the system.
37