Hi Sergey, I copied three linux-6.16-rc2.tar.gz tarball files as the data set. Test Flow: - mkfs on the zram device, mount it - cp tarball files - do idle writeback - check bd_stat writes 185072 pages On Fri, Jun 20, 2025 at 12:15 PM Sergey Senozhatsky <senozhatsky@xxxxxxxxxxxx> wrote: > > Hi, > > On (25/06/20 12:09), Richard Chang wrote: > > Hi Sergey, > > The main idea is to replace submit_bio_wait() to submit_bio(), remove > > the one-by-one IO, and rely on the underlying backing device to handle > > the asynchronous IO requests. > > From my testing results on Android, the idle writeback speed increased 27%. > > > > idle writeback for 185072 4k-pages (~723 MiB) > > $ echo all > /sys/block/zram0/idle > > $ time echo idle > /sys/block/zram0/writeback > > > > Async writeback: > > 0m02.49s real 0m00.00s user 0m01.19s system > > 0m02.32s real 0m00.00s user 0m00.89s system > > 0m02.35s real 0m00.00s user 0m00.93s system > > 0m02.29s real 0m00.00s user 0m00.88s system > > > > Sync writeback: > > 0m03.09s real 0m00.00s user 0m01.07s system > > 0m03.18s real 0m00.00s user 0m01.12s system > > 0m03.47s real 0m00.00s user 0m01.16s system > > 0m03.36s real 0m00.00s user 0m01.27s system > > Has this been tested on exactly same data sets? page-to-page comparable? > We decompress before writeback, so if the data had different compression > ratios, different number of incompressible objects, etc. then the results > are not directly comparable.
