34.2 Throughput

We can also ask what the throughput of our pipeline is. Unlike latency, we will expect throughput to increase considerably through pipelining.

If we execute a task with latency LL (i.e. the task takes LL units of time to complete) then the throughput is 1L\frac{1}{L} - i.e. we will execute that many tasks per unit of time.

Theorem 34.2.1

Let P=(t1,,tn)P=(t_{1},...,t_{n}) be a pipeline where |t1|=|t2|==|tn|=1|t_{1}|=|t_{2}|=...=|t_{n}|=1, then the total throughput of this pipeline is approximately nn.