Alexander Motin

> or call master version 2.99 @clhedrick No. Then the version number would be bigger than later released 2.3, which is not right. Good side of 2.2.99 is that it...

@gbonfiglio I haven't looked too deep when exactly it happened, but we've noticed that introduction of MGLRU eviction mechanisms in kernel changed behavior a lot, though there could be other...

ZFS verifies checksums per block. If the block is not trivially small, it may be difficult to say where the error actually happened. PS: Rather than running scrubs in a...

> Do you mean performing some manual "specialization" on `cityhash4` by replacing some variables with 0? Yes. > I think if we can make `cityhash4` a fully inline function, maybe...

> The question is now whether we should inline the hash function everywhere (code size will slightly grow, but the compiler can always do its best), or provide and export...

> inlining calls to a function will always lead to less (or at least equal) stack usage than caller + callee combined It may be true in some cases due...

> If so, I can just create and export specialized `cityhash1` and `cityhash2`. I would not object. And may be `cityhash3` also.

I wonder why is the difference between city2 and city4 is so insignificant. It would be good to verify that inlining optimization does what is expected, now that all the...

Yea. At least building with Clang on FreeBSD `dnode_hash_city2()` and `dnode_hash_city4()` appeared identical until addition of `__noinline` to `cityhash*()` fixed the optimization (AMD EPYC 7402): ``` mav@srv:~/10% ./hash_bench [==========] Running...

> When change all disk in pool raid5sas to write back i gat write speed pool about 100-150MB/s. Still much to low that maximal write ~450MB/s What are those write-through...