Pipeline Loops – Bottleneck Example
}Buffer Performance
}Assume: FL = 2; BL = 4
}Ring Architecture
}Three tokens with four
full-buffers
}N.B. # of tokens in ring
is constant
}Performance Analysis
}Each token can move
forward every 3 * BL = 12 time units
}Example: token #3 moves forward at t = 2 and t = 14
}Each token completes an
iteration every 12 * 4 = 48 time units
}Three tokens do this concurrently
}Completing 3 iterations every 48 time units
}Yields, throughput of 1
iteration every 48/3 = 16 time units
CE-653 - Asynchronous Pipeline
Performance
11/5/2014
24