Dask dataframes known_divisions and performance

I have several files whose with a column called idx and I would like to use it as index. The dataframe obtained has about 13M row. I know that I can read and assign index in this way (which is slow ~40 s)

df = dd.read_parquet("file-*.parq")
df = df.set_index("idx")

or in this other way (which is quick ~40 ms)

df = dd.read_parquet("file-*.parq", index = "idx")

A simple operation as calculate the length is ~4x faster with the second method. What I don't understand is

• on the first case df.known_divisions returns True while on the second is False. I expected the opposite behaviour. I then did several operations on top of df and the without a known_division I'm always obtaining better performance. I'm scratching my head to figure out if this is happening on purpose or not.
• the number of partitions is the number of files. How can I set a different number of partitions?

UPDATE It is not just calculating len which is faster. In my calculation I create 4 new dataframes using groupby, apply and join several times and these are the timings

|                  |Load and reindex (s)|Load with index (s)|
|:-----------------|-------------------:|------------------:|
| load             |            12.5000 |            0.0124 |
| grp, apply, join |            11.4000 |            6.2700 |
| compute()        |           146.0000 |          125.0000 |
| TOTAL            |           169.9000 |          131.2820 |