Libslock

This repository provides:

• libslock, a cross-platform interface to atomic operations and other common operations
• implementations of a number of well-known locking algorithms
• benchmarks testing the performance of atomic operations and various locking algorithms

The package has been tested on x86_64 Intel and AMD machines, Tilera and Sparc architectures.

A version of this code was used in the paper "Everything you always wanted to know about synchronization but were afraid to ask" (accessible here: http://dl.acm.org/citation.cfm?doid=2517349.2522714).

Makefile parameters:

Locking Algorithm

Can be passed using LOCK_VERSION to the Makefile. LOCK_VERSION can take one of the following values:

• USE_TTAS_LOCKS - use test-and-test-and-set locks
• USE_SPINLOCK_LOCKS - use test-and-set spinlocks
• USE_TICKET_LOCKS - use ticket locks
• USE_HTICKET_LOCKS - use hierarchical ticket locks
• USE_MCS_LOCKS - use MCS locks
• USE_CLH_LOCKS - use CLH locks
• USE_HCLH_LOCKS - use HCLH locks
• USE_ARRAY_LOCKS - use array locks
• USE_RW_LOCKS - use read-write locks (not used in paper, not optimized)
• USE_MUTEX_LOCKS - use the phtread mutex

Platform

Can be passed using PLATFORM to the Makefile; the settings are specific to the platforms we were using (topology, etc.); for other platforms the characteristics can be defined in platform_defs.h. The pre-defined platforms are:

• XEON - 8 x 10-core Intel sever
• OPTERON - 8 x 6-core AMD server
• NIGARA - 8-core SparcT2 machine
• TILERA - 36-core Tilera machine

Detailed descriptions of these platforms can be found in the paper.

The OPTERON_OPTIMIZE option uses some of the Opteron-specific optimizations mentioned in the paper. Atomic operation to be tested

For the benchmarks testing atomic operations, this parameter selects the desired operation. Can be passed to the Makefile using PRIMITIVE:

• TEST_FAI - fetch-and-increment
• TEST_TAS - test-and-set
• TEST_CAS - compare-and-swap
• TEST_SWAP - atomic swap
• TEST_CAS_FAI - fetch-and-increment implemented using compare-and-swap

ALTERNATE_SOCKETS is used for thread placement on the Niagara; if not set, hardware threads begin by being assinged to the same core; if set threads are disitributed evenly among the cores