Fix GpuShuffleCoalesce op time metric doesn't include concat batch time

[res-life]

Fixes #5891

The problem in the code is that the optime metric does not cover the concatenation time performed in the next() method.
Update the range of optime to include the next().

Signed-off-by: Chong Gao res_life@163.com

[res-life]

build

[res-life]

Now the op time covers the concat time:

GpuShuffleCoalesce

concat batch time total (min, med, max )
8 ms (2 ms, 2 ms, 2 ms )
......
op time total (min, med, max )
1 ms (0 ms, 0 ms, 1 ms )

==>>

GpuShuffleCoalesce

concat batch time total (min, med, max )
7 ms (1 ms, 1 ms, 1 ms )
op time total (min, med, max )
24 ms (4 ms, 6 ms, 7 ms )

[res-life]

The op time now also covers the SEMAPHORE_WAIT_TIME time.

[revans2]

The op time now also covers the SEMAPHORE_WAIT_TIME time.

We do not want this. The semaphore wait time is supposed to be a separate thing. Ideally op time is only the time on the GPU.

From https://github.com/NVIDIA/spark-rapids/blob/branch-22.08/docs/tuning-guide.md#metrics

opTime	op time	Time that an operator takes, exclusive of the time for executing or fetching results from child operators, and typically outside of the time it takes to acquire the GPU semaphore.

[revans2]

concat batch time was removed because we are concating the batch on the CPU, not the GPU. opTime was intended to convey the GPU time. I am fine if we redefine it to include CPU time, so long as it is documented clearly and there is a way to pull it back apart again, i.e. we know that the concat time happened on the CPU so the GPU time is opTime - concatTime

[res-life]

build

[res-life]

build