Performance gap on particular lattices

[DotanGazith]

In my research in astronomy, I use g6k to sample many short vectors in particular lattices I define.
The way I do so is by doing BKZ->pump-up of the lattice, and in most instances, I generate from my data, the performance is great and roughly matches the measurements from the 2019 G6K paper.
However, for some specific lattices this process breaks at some late stage of the pump-up, to the point that a sieve instruction in dimension ~80 requires ~1h.

I attach an example lattice that recreates this odd behaviour, specifically, here I used:

• BKZ with dimension 45
• The bgj1 sieving algorithm
• Context going from (0, 43, 88) -> (0, 8, 88) in the (EL, S)*(43-8), with the run-time significantly increasing around (0, 14, 88)

Is there some common wisdom on the way to "amortize" the runtime of such an operation?

slow_lattice.txt