Dask Futures performance as the number of tasks increases

I have been running a performance comparison of Dask Futures and other solutions and found unexpected behavior when it comes to the throughput as I increase the number of tasks.

In this test, I am running on 6 nodes (client, scheduler and 4 workers, 1 thread per worker); however, the behavior is consistent for 2-2048 workers. Each task is a dummy function returning its argument. I am timing this:

```python
futures = client.map(dummy, data, key=labels)
wait(futures,return_when='ALL_COMPLETED')
```
where `data` is a list of `None`, `labels` is a unique key, and the `dummy` function simply returns its argument. Complete code is attached.

Based on 30 repetitions for each configuration and after discarding the first timing to account for any lazy initialization, I am getting the following results:


Is this characteristic something you would expect given such a trivial task? Is the scheduler not able to keep up with this many tasks?

Thank you for any insights!

**Environment**:
- dask==2021.12.0, dask_mpi==2021.11.0, distributed==2021.12.0
- Python 3.10.1
- Cray XC-40, each node has 32 cores and 128 GB of RAM
- reproduced on another cluster with dask 2021.11

**Complete code**:
[latency.py.txt](https://github.com/dask/distributed/files/7949342/latency.py.txt)