Besides the disability of global congestion prediction, the
practice above only handles average delays in NoC, but
rarely on worst-case. Recently, approaches based on network
calculus [7][8] have been proposed to derive delay bound for
application-specific NoCs [9]. Results have been obtained on
wormhole routing NoCs [10] and credit based flow control
router [11]. But congestion prevention problems are rarely
discussed. This research gap motivates us to analyze the worstcase
delay bound and globally balance the on-chip traffic
workload to reduce delay both in average and maximum case.