NB-FEB: A Universal Scalable Easy-to-Use Synchronization Primitive for Manycore Architectures

NB-FEB: A Universal Scalable Easy-to-Use Synchronization Primitive forManycore Architectures

Phuong H. Ha (Univ. of Tromsø, Norway) Philippas Tsigas (Chalmers Univ., Sweden)Otto J. Anshus (Univ. of Tromsø, Norway)

Presentation at OPODIS ’09

December 15-18, 2009, Nimes, France

Problem Manycores require scalable strong synchronization

primitives. Conventional strong primitives do not scale well enough

for manycores [UCB Landscape]. Contention on a synchronization variable increases with

the number of processing cores.

OPODIS '09, Nimes, France

2 cores 16 cores 1000 cores

Desired features New synch. primitives for manycores should

be: Scalable

1000s of cores

Universal powerful enough to support any kind of synchronization

(like CAS, LL/SC)

Feasible able to implement in hardware

Easy-to-use


Our main contributions

A novel synch. primitve with all these features


• Non-blocking Full/Empty Bit (NB-FEB)• NBFEB-STM: a non-blocking STM

Road-map

NB-FEB Feasible Universal Scalable Easy-to-use

NBFEB-STM: a non-blocking STM


Feasibility Key idea: slight modifications of a widely deployed primitive

A variant of the original FEB that always returns a value instead of waiting for a conditional flag


Test-Flag-and-SetTFAS( x, v) { (o, flago) (x, flagx); if flagx = false then (x, flagx) (v, true); end if return (o, flago);}

Store-And-ClearSAC( x, v) { (o, flago) (x, flagx); (x, flagx) (v, false); return (o, flago);}Store-And-SetSAS( x, v) { (o, flago) (x, flagx); (x, flagx) (v, true); return (o, flago);}

LoadLoad( x) { return (x, flagx);}

Original FEB: Store-if-Clear-and-SetSICAS(x,v) { Wait for flagx to be false; (x, flagx) (v, true);}

Universality Key idea: write-once objects with 3+ states

TFAS Wait-free consensus, n


Decision (, false);

TFAS_Consensus( proposal) { (first, ) TFAS(Decision, proposal); if first = then return proposal; else return first;}

Scalability Key idea: Combinability

eliminates contention & reduce load Ex: TFAS


x=

TFAS(x,1)TFAS(x,2)

TFAS(x,3)TFAS(x,4)

TFAS(x,1) TFAS(x,3)

TFAS(x,1)

x=1

1 1 1

1

TFAS( var x, value v) atomically { (o, flago) (x, flagx); if flagx = false then (x, flagx) (v, true); end if return (o, flago);}

Note: CAS or LL/SC is not combinable

NB-FEB combining logic


(x, [v1]) Successive primitive with parameter (x, [v2])

Load SAC SAS TFAS

Load Load SAC(v2) SAS(v2) TFAS(v2)

SAC SAC(v1) SAC(v2) SAS(v2) SAS(v2)

SAS SAS(v1) SAC(v2) SAS(v2) SAS(v1)

TFAS TFAS(v1) SAC(v2) SAS(v2) TFAS(v1)

Easy-to-use

Key idea: abstractions for productivity-layer programmers Non-blocking software transactional memory NBFEB-STM


Road-map

NB-FEB Feasible Universal Scalable Easy-to-use

NBFEB-STM: a non-blocking STM


NBFEB-STM

Models Objects are accessed within transactions No nested transactions Garbage collected programming languages (e.g.

Java)

Features Obstruction-free STM Eliminate conventional synch. hot spots in STMs Optimal space complexity (N)


Challenge 1: TFAS-SAC interleavingCAS-based STMs NBFEB-STM

Need SAC to clear pointer’s flag

Overlapping TFAS1 & TFAS2 both may succeed due to SAC’s interference.

violate TMObj’s semantics


TMObj

Old NewTM0

Old NewTM1

Copy

CAS1

Old NewTM2

CopyCAS2

locator

Key idea 1 Keep a linked list of locators

write-once pointer next


locator

Old NewTM0

Old NewTM1

Old NewTM2

eliminate SAC interference

TFAS

TFAS

Challenge 2: Space complexity


CAS-based STMs NBFEB-STM

TMObj

Old NewTM0

Old NewTM1

Old NewTM2

CAS2

locator locator

Old NewTM0

Old NewTM1

Old NewTM2

Key idea 2 Only the head is needed for further accesses break the list of obsolete locators


locator

Old NewTM0

Old NewTM1

Old NewTM2

pi

pi

SAC

Optimal space complexity (N)

Challenge 3: Find the head


locator

Old NewTM0

Old NewTM1

Old NewTM2

pi

pi

HeadX

Key idea 3 No nested transactions one active locator / thread


i

0

1

……

N

TMObj locator

Old NewTM0

Old NewTM1

Old NewTM2

pi

pi

SAC

Correctness

NBFEB-STM fulfills the essential aspects of TM [Guerraoui, PPoPP ’08] Instantaneous commit Precluding inconsistent views Preserving real-time order


Conclusions

Introduce a novel non-blocking full/empty bit primitive (NB-FEB) Scalable, universal, feasible and easy-to-use

Provide an abstraction, NBFEB-STM, built on top of the primitive.



Thanks for your attention!

NB-FEB: A Universal Scalable Easy-to-Use Synchronization Primitive for Manycore Architectures

Documents

flagx x

flago x

return x

flagx v

false x

loadload x

parameter x

settfas x