Tuning linux's load balancing algorithm for CMT system

摘要

Linux 2.6 load balancing algorithm on scheduling domain supports CMP, CMT, SMP, NUMA architecture. For SMT, the algorithm try to assign the new process to the idlest CPU of the idlest core, and if the first CPU of a core is comparati vely idle, it tries to pull moderate amount of tasks from the busiest CPU of the core to balance the system workload periodically. Under certain circumstances, this strategy would cause the system to be more unbalanced. For the above defects, the algorithm can be adjusted in two ways. (1) The idlest CPU of the entire system should be selects for the new process; (2) the idlest CPU of a scheduling domain periodically can move tasks from the scheduling domain to itself. After applying this optimization, the system's performance achieved the highest increase of 8% in hackbench testing on an 8-core 32-thread XLR532 processor.