my earlier tests showed that in some disk-bound workloads smaller pbxt_checkpoint_frequency gives less fsyncs. For example pbxt_checkpoint_frequency=100MB gave 2x fsyncs as pbxt_checkpoint_frequency=1MB. My test case was not exactly like this and data set was smaller but I think at the engine level there could be similarities. Of course this is not the solution of the entire problem but this can improve performance...
my earlier tests showed that in some disk-bound workloads smaller pbxt_checkpoint _frequency gives less fsyncs. For example pbxt_checkpoint _frequency= 100MB gave 2x fsyncs as pbxt_checkpoint _frequency= 1MB. My test case was not exactly like this and data set was smaller but I think at the engine level there could be similarities. Of course this is not the solution of the entire problem but this can improve performance...