A loop partition technique for reducing cache bank conflict in multithreaded architecture

Parallel multithreaded architectures take advantage of the ability to execute more than one thread simultaneously on a single chip at low synchronisation and communication costs and high hardware resource utilisation. However, a high bandwidth cache, such as a multibank cache, is especially critical to serve memory accesses issued at the same time from different threads. To prevent bank conflicts of multibank cache from seriously degrading system performance, a loop partition method is proposed to reduce or even eliminate bank conflicts. The partition allows each thread access to certain bank modules and prevents any two from accessing the same bank module. The method neither slows down the clock rate nor increases the array subscript expression complexity. The performance gains of the bank-conflict-free loop partition approach are shown in simulation results.