RCU: Reader-Writer Locking and Garbage Collection for Parallel Processing, Study notes of Computer Science

Download RCU: Reader-Writer Locking and Garbage Collection for Parallel Processing and more Study notes Computer Science in PDF only on Docsity!

Outline

• Intro – Phil• Modern Processors – why – Jim • RCU – what – Phil • Hash Tables – how - Josh

Synchronization

Read Copy Update (RCU)

A

B

C

D

C’

RCU – What does this buy us? • Readers need

NO^

synchronization

• Single writer in parallel with readers

Need for Garbage Collection

A

B

C

D

C’

RCU – What does this cost us? • Need for garbage collection• Potential for “stale data”

RCU Garbage Collection

Linux Style

• RCU is part of a system call• system calls don’t save state betweencalls• Grace Periods can be defined in terms ofcontext switches^ – no pre-emption– every processor goes through a contextswitch • No need to keep track of individualprocesses/threads

Read-side APIs RWL

-^ read_lock()

-^ read_unlock()

RCU

-^ rcu_read_lock()

-^ rcu_dereference()

-^ rcu_read_unlock()

Linearization Points

• A single point in time where an actiontakes effect• Correctness Criteria• Allows actions to be treated as atomic

Write-side Linearization Points

RWL

-^ write_lock()

-^ write_unlock()

RCU

-^ lock()

-^ rcu_assign_pointer()

-^ synchronize_rcu()

-^ free()

-^ unlock()

Which dereference?^ A

B

C

D

C’

RWL Stale Data

request_read_lock()acquire_read_lock()

computationinternal actionread_unlock()

external_action

RWL Stale Data

request_read_lock()acquire_read_lock()

request_write_lock()

acquire_write_lock()

update_data

computation

write_unlock()

internal actionread_unlock()

external_action

RCU Stale data

rcu_read_lock()rcu_dereference()

computationinternal actionrcu_read_unlock()

external_action