Algorithms for scalable synchronization on shared-memory multiprocessors

Abstract

Busy-wait techniques are heavily used for mutual exclusion and barrier synchroni?;ation in shared-memory parallel programs. Cnfortunatcly, typical implementations of busy-waiting tend to produce large aIIlounts of memory and interconnect contention, introducing perf(H"-IIl<-1IlCe bottlenecks that become markedly more pronounced <-1,, -: applications scale. \:Ve argue in this paper that this problem is not fundamental, and tha.t one can in fad construct bUi:>Y-wait synchroni7';ation algorithms that induce no memory or interconnect contention. The key to these algorithms is for every processor to spin on a separate location in local memory, and for some other processor to terminate the spin \vith a single remote \vrite operation at an appropriate time. Locations on \vhich to spin may be local as a result of coherent caching, or by virtue of static allocation in the local portion of phYi:>ically dii:>tributed i:>hared memory.'Ve present a new scalable algorithm for spin locks that generates 0(1) remote references per lock acquisition, independent of the number of processors attempting to acquire the lock. Our algorithm provides rea .. .;;onable latency in the absence of contention, requires only a constant amount of i:>pace per lock, and requirei:> no harchvare support other than a i:>\vap-with-memory instruction. 'ile ali:>o prei:>ent a nev, ,! i:>calable barrier algorithm that generates 0(1) remote references per processor reaching the barrier, and observe that two previouslyknmvn barriers can likewise be cast in a form that spins only on local locations. Kone of these barrier algorithms requires hardware support beyond the usual atomicity of memory reads and \vrites.'ile compare the performance of our scalable algorithms with other software approaches to busy-wait synchroniz;ation on both a Sequent Symmetry and a BBN Butterfly. Our principal conclusion is that contention due to s;qnchronization need not be a problem in large-scale 8han;rl-meTrwTY mnltiprnce88oT8. The existence of scalable algorithms greatly weakens the ca.':>e for costly special-purpose harchvare i:>upport for i:>ynchronization, and provides a case against so-called "dance hall" architecturei:>, in which shared memory locations are equally far from all processors.

The Black-White Bakery Algorithm and Related Bounded-Space, Adaptive, Local-Spinning and FIFO Algorithms

“…Several interesting mutual exclusion algorithms have been published in recent years that are either adaptive to contention or satisfy the local-spinning property [3,4,6,7,9,10,14,21,24,34,37,41,44]. (These two important properties are defined in the sequel.)…”

“…Three important papers which have investigated local-spinning are [9,21,34]. The various algorithms presented in these papers use strong synchronization primitives (i.e., stronger than atomic registers), and require only a constant number of remote accesses for each access to a critical section.…”

Pessimistic Software Lock-Elision

“…This is state-of-the-art RW Lock implementation for Intel platform. MCS-Lock Michael and Scott's MCS Lock [16].…”

Dense Gaussian Networks: Suitable Topologies for On-Chip Multiprocessors

“…Broadcast communications are employed in many parallel applications such as matrix multiplication, LU factorization, Householder transformations and other basic linear algebra algorithms. Moreover, important architectural issues such as maintaining cache coherency and supporting barrier synchronization in multiprocessors may depend on the ability of the network to perform broadcasting communication (11,20) .…”