A substantial regression in latency can be observed for SPECjbb2015 using G1.

Benchmark................Linux aarch64.......Linux x64
SPECjbb2015 (G1)
     critical10ms................(*)..-5.11%.....(*)..-8.78%
     criticalJOPS................(*)..-3.10%....(*)..-2.50%
     maxJOPS.........................-0.12%...........0.20%

(*) = statistical signficance
30 runs of each baseline (23) and target (25)

I have ruled out NUMA and large pages as a reason for the regression.

Currently investigating which particular build between 23 and 25 introduced the performance regression.