"I've got two workloads which use the same device, one randomly the other serially. The disk service times will be very different, but I only get one average service time from the measurements. How can I represent this in a model?"

Let's pick some nice easy numbers (to make the calculation simple, but the technique works even with awkward ones). Workload 1 does 20 I/Os per second reading serially. Workload 2 does 5 per second reading randomly. You estimate that each serial I/O will take 5ms and each random one 15ms. What you actually get from the measurements is something like 7 because it's an average, and that's a useless figure for either workload. What to do?

Set the service time of the device to 5ms. This will be fine for the serial-read component, but that's too low for the random one by a factor of three. Simply multiply the I/O rate for the random component by three.

Applying that (wrong) service time in a model will give:
Workload 1 - (20 x 7) per second = 140 ms in a second - 14% utilization -
too high Workload 2 - (5 x 7) per second - 35 ms in a second - 3.5%
utilization - too low.

Now do the multiplication:
Workload 1 - (20 x 5) per second - 100ms in a second - 10% device
utilization - OK
Workload 2 - (15 x 5) per second - 75 ms in a second - 7.5% utilization - OK

Both workloads are now right, and the total device utilization comes out right too.

Motto - Always look both ways before crossing the road.