Nonblocking Algorithms and Preemption-Safe Locking on Multiprogrammed Shared Memory Multiprocessors

Abstract

Most multiprocessors are multiprogrammed in order to achieve acceptable response time and to increase their utilization. Unfortunately, inopportune preemption may significantly degrade the performance of synchronized parallel applications. To address this problem, researchers have developed two principal strategies for concurrent, atomic update of shared data structures: (1) preemption-safe locking and (2) non-blocking (lock-free) algorithms. Preemption-safe locking requires kernel support. Non-blocking algorithms generally require a universal atomic primitive such as compare-and-swap or load-linked/store-conditional, and are widely regarded as inefficient.We evaluate the performance of preemption-safe lock-based and non-blocking implementations of important data structures-queues, stacks, heaps, and counters-including non-blocking and lock-based queue algorithms of our own, in micro-benchmarks and real applications on a 12-processor SGI Challenge multiprocessor. Our results indicate that our non-blocking queue consistently outperforms the best known alternatives, and that data-structure-specific non-blocking algorithms, which exist for queues, stacks, and counters, can work extremely well. Not only do they outperform preemption-safe lock-based algorithms on multiprogrammed machines, they also outperform ordinary locks on dedicated machines. At the same time, since general-purpose non-blocking techniques do not yet appear to be practical, preemption-safe locks remain the preferred alternative for complex data structures: they outperform conventional locks by significant margins on multiprogrammed systems.

The Repeat Offender Problem: A Mechanism for Supporting Dynamic-Sized, Lock-Free Data Structures

“…Performance experiments presented in that paper show that the overhead of the dynamic-sized FIFO queue over the non-dynamic-sized one of [11] is negligible in the absence of contention, and low in all cases.…”

“…In [6] we present a dynamic-sized lock-free FIFO queue achieved by applying our ROP solution (without the optimizations discussed above) to the non-dynamic-sized implementation of Michael and Scott [11] together with the non-dynamic-sized freelist of Treiber [13]. Performance experiments presented in that paper show that the overhead of the dynamic-sized FIFO queue over the non-dynamic-sized one of [11] is negligible in the absence of contention, and low in all cases.…”

“…Another common approach is to augment values in objects with version numbers or tags, and to access such values only through the use of compare-and-swap (CAS), such that if a CAS executes on an object after it has been deallocated, the value of the version number or tag will ensure that the CAS fails [13,11,12]. In this case, the version number or tag value must be carried with the object through deallocation and reallocation.…”

“…In this case, the version number or tag value must be carried with the object through deallocation and reallocation. A common approach for achieving this is to maintain explicit "freelists", which contain objects that are not currently in use [13,11]. (In the remainder of the paper, we refer to freelists as "object pools" in order to avoid confusing "freeing" an object-by which we mean returning it to the memory allocator through the free library routine-with placing an object on a freelist.)…”

Preemption Adaptivity in Time-Published Queue-Based Spin Locks

“…Unfortunately, while excellent nonblocking implementations exist for many important data structures (only a few of which we have room to cite here [21,23,24,29,30]), general-purpose mechanisms remain elusive. Several groups (including our own) are working on this topic [6,10,18,26], but it still seems unlikely that nonblocking synchronization will displace locks entirely soon.…”