Abstract: Techniques are disclosed to provide scalable spinlocks for non-uniform memory access (NUMA). In some examples, a global spinlock configured to protect access to a shareable resource is protected by multiple local spinlocks, which are each configured to control access to the global spinlock. In a multi-socket NUMA system, the global spinlock is allocated on one of the sockets, and the local spinlocks are distributed over the multiple sockets. In some embodiments, one local spinlock is allocated on each of the multiple sockets. In other embodiments, the multiple local spinlocks may be equally distributed over the NUMA sockets. When contention for the global spinlock is low, processes can attempt to directly acquire the global spinlock. In contrast, when contention for the global spinlock is high, processes need to first acquire one of the local spinlocks associated with the global spinlock before attempting to acquire the global spinlock.

Abstract: Techniques are disclosed to provide scalable spinlocks for non-uniform memory access (NUMA). In some examples, a global spinlock configured to protect access to a shareable resource is protected by multiple local spinlocks, which are each configured to control access to the global spinlock. In a multi-socket NUMA system, the global spinlock is allocated on one of the sockets, and the local spinlocks are distributed over the multiple sockets. In some embodiments, one local spinlock is allocated on each of the multiple sockets. In other embodiments, the multiple local spinlocks may be equally distributed over the NUMA sockets. When contention for the global spinlock is low, processes can attempt to directly acquire the global spinlock. In contrast, when contention for the global spinlock is high, processes need to first acquire one of the local spinlocks associated with the global spinlock before attempting to acquire the global spinlock.