Well, on one of my computer, WU calculation length goes from 1h to 12h:

https://docs.google.com/spreadsheets/d/1O4lnYzuh7xHEFKyyKzAIDyyHafwLOz-8FjjxxL5ZjrA/edit?usp=sharing

Computer: https://numberfields.asu.edu/NumberFields/show_host_detail.php?hostid=2349835

Is there a reason for this ?
(it happens to all of my machines, this one is just an exemple as I happen to have >300 tasks pending to send)

Well, on one of my computer, WU calculation length goes from 1h to 12h:

https://docs.google.com/spreadsheets/d/1O4lnYzuh7xHEFKyyKzAIDyyHafwLOz-8FjjxxL5ZjrA/edit?usp=sharing

Computer: https://numberfields.asu.edu/NumberFields/show_host_detail.php?hostid=2349835

Is there a reason for this ?
(it happens to all of my machines, this one is just an exemple as I happen to have >300 tasks pending to send)

On a 4GHz single treaded core I see times ranging from .3 hrs to 4 hrs with an average of 1.8 hrs (very rarely there are outliers that are only a few seconds, but I will ignore those).
The ratio of slowest to fastest is ~13, which is in line with your observation.

The quick answer is yes, there is a well understood reason for it. The smallest unit in a dataset is a congruence vector. This latest case (14x271) has 19,660,800 Cvecs total. I decided to break the case into 3,932,160 WUs with 5 Cvecs per WU. The time for a single Cvec has high variance, so the more Cvecs per WU the more uniform the time. I try not to go below 5 Cvecs per WU, otherwise the variance is too high (as you are seeing here). On the flip side, increasing the Cvecs per WU reduces the variance, but the average time goes up.