by

Effective locking is necessary for satisfactory performance on large Itanium based NUMA systems. Synchronization of parallel executing streams on NUMA machines is currently realized in the Linux kernel through a variety of mechanisms which include atomic operations, locking and ordering of memory accesses. Various synchronization methods may also be combined in order to increase performance. The talk presents the realization of basic synchronization in Linux on Itanium and then investigates more complex locking schemes. The current Linux locking mechanisms rely heavily on a simple spinlock implementation that may be fitting for systems of up to 8 processors. However, spinlocks cause excessive cache line bouncing if more processors are contending for a lock. Some approaches that have so far been made to solve the contention issue are presented and it is then suggested to use an implementation for Linux of the approach first proposed by Zoran Radovic which he called "Hierarchical Backoff Locks"