Cooperative Task Management without Manual Stack Management or, Even-driven Programming is Not the Opposite of Threaded Programming A. Adya, J. Howell, M. Theimer, W. Bolosky and J. Douceur

Presented by Will Landecker

1

Outline ●

Introduction: threads vs events?

●

Task management

●

Stack management

●

Hybrid stack management

●

Conclusions

2

Threads vs. events? Events

Threads

●

Explicit yielding

●

Preemptive scheduling

●

Asynchronous I/O

●

Synchronous I/O

–

●

No blocking

–

Continuations

●

3

Blocking

Save state, resume automatically

Threads vs. events? Events

Threads

●

Explicit yielding

●

Preemptive scheduling

●

Asynchronous I/O

●

Synchronous I/O

–

●

No blocking

–

Continuations

●

4

Blocking

Save state, resume automatically

Threads vs. events? Events

Threads

Task management ●

Explicit yielding

●

Preemptive scheduling

●

Asynchronous I/O

●

Synchronous I/O

–

●

No blocking

–

Continuations

●

Blocking

Save state, resume automatically

Stack management 5

Outline ●

Introduction: threads vs events?

●

Task management

●

Stack management

●

Hybrid stack management

●

Conclusions

6

Task management ●

A task encapsulates a control flow

●

Preemptive A

●

time B

A

C

Serial

time A

●

...

B

B

C

Cooperative

...

time B 7

YIELD!

A

C

B

...

Task management ●

A task encapsulates a control flow

●

Preemptive – – –

●

Processes interleave / overlap on uni/multiprocessor Shared state, need synchronization / protection Synchronous I/O

Serial

time A

●

B

C

Cooperative

...

time B 8

YIELD!

A

C

B

...

Task management ●

A task encapsulates a control flow

●

Preemptive – – –

●

Serial – – –

●

Processes interleave / overlap on uni/multiprocessor Shared state, need synchronization / protection Synchronous I/O Task runs to completion Easier to guarantee integrity of data I/O must be asynchronous

Cooperative

time B 9

YIELD!

A

C

B

...

Task management ●

A task encapsulates a control flow

●

Preemptive – – –

●

Serial – – –

●

Processes interleave / overlap on uni/multiprocessor Shared state, need synchronization / protection Synchronous I/O Task runs to completion Easier to guarantee integrity of data I/O must be asynchronous

Cooperative – –

Explicit yield: similar guarantees about data integrity I/O must still be asynchronous 10

Outline ●

Introduction: threads vs events?

●

Task management

●

Stack management

●

Hybrid stack management

●

Conclusions

11

Stack management ●

Manual stack management – – – –

●

Rip tasks into pseudo-atomic event handlers Explicitly save task state to heap Messages, continuations Code becomes difficult to maintain and read as software evolves

Automatic stack management – – – –

1 task 1 procedure Can ignore calls to I/O Control returned automatically Maintainable, readable code 12

Outline ●

Introduction: threads vs events?

●

Task management

●

Stack management

●

Hybrid stack management

●

Conclusions

13

automatic

“sweet spot”

manual

Stack management

Hybrid stack management

event-driven

multithreaded

cooperative preemptive Task management

Claim: combine maintainability and readability of automatic stack management with reasoning benefits of cooperative task management. In reality: New system allows both manual and automatic stack management. 14

Hybrid stack management New system allows both manual and automatic stack management. What could go wrong?

15

Hybrid stack management New system allows both manual and automatic stack management. What could go wrong?

A

B

Manual calling automatic … I/O ...

Given a manual stack management procedure A, and an automatic stack management procedure B.

time

A calls B, B blocks on I/O.

B

16

A

Hybrid stack management Main fiber

Fibers (Windows) are scheduled cooperatively.

otherFiber

A

Adaptors switch between fibers. B

Main fiber runs a scheduler. Procedures expecting automatic stack management must run on a different fiber.

C

B

Each code style (manual vs. automatic stack management) is unaware of the other. time

17

… I/O ...

A

Hybrid stack management Automatic calling manual

B

Given a manual stack management procedure A, and an automatic stack management procedure B.

A

… I/O ...

B

B calls A, A requests I/O. A

time

18

B

Hybrid stack management Main fiber

Solution:

Sched.

otherFiber

Execute B and A on a new fiber.

B

A sees asynchronous I/O. Scheduler ensures B only gets one return.

A I/O request

Sched. I/O complete

A B Sched. 19

Outline ●

Introduction: threads vs events?

●

Task management

●

Stack management

●

Hybrid stack management

●

Conclusions

20

●

●

Cooperative task management is good for reasoning about concurrency. Manual stack management can be cumbersome.

automatic

●

Threads vs events? Or stack management vs task management? Stack management

●

“sweet spot”

manual

Conclusions

event-driven

multithreaded

cooperative preemptive Task management

A “hybrid” stack management system uses fibers to allow both manual and automatic stack management. 21