Speed-up and efficiency

 

We have cast our results in terms of the real time ratio, since this is the most important quantity when one wants to get a practical study done. In this section, we will translate our results into numbers of speed-up, efficiency, and scale-up, which allow easier comparison for computing people.

Let us define speed-up as

$S(p)\; :=\; T(1)\; T(p)\; ,$

where $p$ is again the number of CPUs, $T(1)$ is the time for one time-step on one CPU, and $T(p)$ is the time for one time step on $p$ CPUs. Depending on the viewpoint, for $T(1)$ one uses either the running time of the parallel algorithm on a single CPU, or the fastest existing sequential algorithm. Since TRANSIMS has been designed for parallel computing and since there is no sequential simulation with exactly the same properties, $T(1)$ will be the running time of the parallel algorithm on a single CPU. For time-stepped simulations such as used here, the difference is expected to be small.

Now note again that the real time ratio is $rtr(p)\; =\; 1 sec\; /\; T(p)\; .$ Thus, in order to obtain the speed-up from the real time ratio, one has to multiply all real time ratios by $T(1)/(1 sec)$. On a logarithmic scale, a multiplication corresponds to a linear shift. In consequence, speed-up curves can be obtained from our real time ratio curves by shifting the curves up or down so that they start at one.

This also makes it easy to judge if our speed-up is linear or not. For example in Fig. 13 bottom, the curve which starts at 0.5 for 1 CPU should have an RTR of 2 at 4 CPU, an RTR of 8 at 16 CPUs, etc. Downward deviations from this mean sub-linear speed-up. Such deviations are commonly described by another number, called efficiency, and defined as

$E(p)\; :=\; T(1)/p\; T(p)\; .$

Fig. 14 contains an example. Note that this number contains no new information; it is just a re-interpretation. Also note that in our logarithmic plots, $E(p)$ will just be the difference to the diagonal $p\; \hspace{0.17em}\; T(1)$. Efficiency can point out where improvements would be useful.

  
Figure: Efficiency for the same configurations as in Fig. 13 bottom. Note that the curves contain exactly the same information.