Many of these reads will be metadata reads, as your guest VM does not have enough RAM to cache them. So lots of small reads from guest standpoint, which will be pretty much invisible on the bandwidth graph. Your ARC is pretty big though, so likely cached on the host.
I assume your graphs with storjv3 show measurements from the guest, and with starlight from the host?
It should make sense that if the guest is accessing mostly cached data, it’s the CPU that will be the bottleneck. What type of virtual bus/controller are you using?
I suspect if you reduced ARC a bit and increased your VM memory by the same amount, file walkers might go faster due to less host-to-guest copying.
Expired pieces are checked every hour.