Exploiting multi-level scratchpad memories for time-predictable multicore computing.