Ananiev, Leonid I writes:
> >From Leonid Ananiev
Hello,
>
> Moving dirty pages balancing from user to kernel thread pdfludh entirely
> reduces extra long write(2) latencies, increases performance.
>
[...]
> The benchmarks IOzone and Sysbench for file size 50% and 120% of
> RAM size on Pentium4, Xeon, Itanium have reported write and mix
> throughput increasing about 25%. The described Iozone > 0.1 sec write(2)
Results are impressive.
Wouldn't this interfere with current->backing_dev_info logic? This field
is set by __generic_file_aio_write_nolock() and checked by
may_write_to_queue() to force heavy writes to do more pageout. Maybe
pdflush threads should set this field too?
> latencies are deleted. The condition writeback_in_progress() is tested
> earlier now. As a result extra pdflush works are not created and number
> of context switches increasing is inside variation limites.
Nikita.