On Mon 13-01-25 18:31:07, Alexey Zhuravlev wrote: > Hi, > > I benchmarked rhashtable based patch vs Jan's patch: > > records vanilla rhashtable JK patch > 2.5M records 102s 29s 25s > 5.0M records 317s 28s 30s > 6.0M records -- 35s 44s > > the tests were done using 4.18 kernel (I guess this doesn't matter much > in this context), using an SSD. time to mount after a crash (simulated > with read-only device mapper) was measured. unfortunately I wasn't able > to reproduce with more records as my test node has just 32GB RAM, Thanks for performing the tests! So I guess the dynamic sizing of revoke table before replay is indeed good enough. Let me do a proper patch submission. Honza > > > thanks, Alex > > > On Mon, 4 Nov 2024 20:05:50 +1100 > Li Dongyang <dongyangli@xxxxxxx> wrote: > > > Resizable hashtable should improve journal replay time when > > we have million of revoke records. > > Notice that rhashtable is used during replay only, > > as removal with list_del() is less expensive and it's still used > > during regular processing. > > > > before: > > 1048576 records - 95 seconds > > 2097152 records - 580 seconds > > > > after: > > 1048576 records - 2 seconds > > 2097152 records - 3 seconds > > 4194304 records - 7 seconds > > -- Jan Kara <jack@xxxxxxxx> SUSE Labs, CR
