On Mon, Sep 5, 2022 at 10:17 PM Sergey Senozhatsky <senozhatsky@xxxxxxxxxxxx> wrote: > > On (22/09/05 22:06), Barry Song wrote: > > > > make sense! thanks! i assume you will have some benchmark data to compare > > three cases, > > 1. lzo with recompress zstd > > 2. always use lzo > > 3. always use zstd > > > > such as power consumption, cpu utilization, available memory, benefits to user > > experiences especially to UI smoothness under memory pressure? > > So I didn't want to include any benchmarks, because this is entirely > specific to device's data sets/patterns. In term of CPU usage, zstd > decompression is really fast [1]; and the way plan to use is battery > aware - e.g. when low on battery do not recompress at all, if AC is > plugged in then recompress more aggressively, etc. > > In term of benchmarks... a copy paste from our internal tests. But > *do note* that this is relative only to our specific data sets. > Your millage may vary. > > ZSTD recomp algorithm (5.10 kernel, so the last column is the number of > 'zram huge pages'): > > - Initial state of zram swap partition > localhost ~ # cat /sys/block/zram0/mm_stat > 8955662336 2180671776 2277711872 0 3179720704 798724 469474 118949 > > - Recompress HUGE objects only > localhost ~ # echo huge > /sys/block/zram0/recompress > localhost ~ # cat /sys/block/zram0/mm_stat > 8944390144 2106998658 2211835904 0 3179720704 798617 469474 66821 > > - Recompress IDLE pages that are >= 3000 bytes in size > localhost ~ # echo 3000 > /sys/block/zram0/recompress > localhost ~ # cat /sys/block/zram0/mm_stat > 8934166528 2085232505 2207690752 0 3179720704 798484 469474 66811 > > - Recompress the remaining IDLE pages that are >= 2000 bytes in size > localhost ~ # echo 2000 > /sys/block/zram0/recompress > localhost ~ # cat /sys/block/zram0/mm_stat > 8913981440 1946488434 2145157120 0 3179720704 798130 469474 66498 > > - Recompress the remaining IDLE pages that are >= 1000 bytes in size > localhost ~ # echo 1000 > /sys/block/zram0/recompress > localhost ~ # cat /sys/block/zram0/mm_stat > 8905592832 1711533182 1984495616 0 3179720704 797162 469474 66222 > > [1] https://facebook.github.io/zstd/ Thanks very much. I guess the difficulty is that we need to comprehensively evaluate its effect on a real product such as android. there is always a trade-off between cpu utilization and more available memory. Best Regards Barry
