Unfortunately, and amazingly
after all these years of research, there are no reproducible examples in the
literature with absolute timings substantiating the performance of any
vendor's parallel model. Typically, only virtually meaningless figures
of 'parallel speedup' over their own undisclosed serial run times are given. So, we can only present a hypothetical
example, using commonly observed Sensor serial speedups.
Sensitivity of simulation results to 10
variables is examined. A base run is the best guess. Each
variable will be assigned low, best guess, and high values. So 21 runs
need to be made. This or something similar might be done for history
matching, optimization, or uncertainty quantification, possibly for training
of a proxy model or in
computing gradients, or in generation of probabilistic reserves estimates.
One user has a 16-node cluster and a
parallel simulator. He's getting parallel speedup of about 10, let's say
from a 20 hour serial job to a 2 hour parallel job. All runs are complete
in 42 hours.
Another user has Sensor and 11 pc's (or
that same 16-node cluster). Let's say this is a black oil case, and
Sensor is only 3 times faster than the first model, in serial. Sensor
completes all runs in 13.33 hours. If it was a compositional case and
Sensor was 8 times faster, Sensor would complete all runs in 5 hours.
If the Sensor user had the
first user's cluster, he could look at sensitivities to 15 variables instead
of 10, in the same time (this would take 62 hours for the parallel user).
The same 21 runs are to be made. The
parallel user has five 16-node clusters. So each cluster must run 4 jobs,
and one cluster must run another one. All runs are complete in 10 hours.
The Sensor user has 21 pc's,
representing only about a quarter of the total computing power of the
parallel user. For the black oil case, Sensor completes all runs in 6.67 hours.
For the compositional case, in which Sensor has a greater serial speed
advantage, all runs are complete in 2.5 hours. If the Sensor user had
those five clusters, he could look at sensitivities to 39 variables instead
of 10 in the same time (this would take the parallel user 32 hours).
Let's say that instead of 21 runs, 99
runs are needed (for 49 variables).
The parallel user needs to make 20 sets
of runs on his five clusters. All runs are complete in 40 hours.
The Sensor user has 50 pc's. For the black oil case, all runs are
complete in 13.33 hours. For the compositional case, Sensor completes
all runs in 5 hours. The Sensor user could look at sensitivities to 79
variables instead of 49 in the same time, if he had the five clusters (this
would take 64 hours for the parallel user).
Please check our numbers by comparing
model performance on individual runs.