[zfs-discuss] L2ARC and SLOG on HW RAID with writeback cache
g.danti at assyoma.it
Sun Apr 29 10:13:01 EDT 2018
Original OP here...
Il 29-04-2018 15:28 Richard Yao via zfs-discuss ha scritto:
> I have seen RAID 1 randomly serve the wrong information because one of
> the copies was wrong. There was nothing to “perform the check”. A
> misdirected write will happily cause the wrong data to be present.
> Even if you are lucky and the sector(s) that get clobbered are
> clobbered sin such a way that it fails internal ECC, the sectors where
> the data was supposed to go still contain the wrong data.
SAS + T10 provide complete support against bitroot / misdirected writes.
However this is not my expected usage: I'm planning to use plain SATA
disks, so ZFS checksum is *really* important to me (I also had multiple
SATA data corruption in the past, mainly due to bad power supply).
In fact, I did not ask if using ZFS vs HWRAID for the main pool. Rather
I was laser-focused on L2ARC and SLOG, because:
- corrupted L2ARC entries can be transparently reloaded from main pool;
- corrupted SLOG entries shoule be *really* rare, as data reside in ZIL
for some seconds at most *and* are only read after an host crash.
In current systems, which I build around IT or integrated controller, I
use a small MD RAID1 array for OS installation and some non-raid,
separate partitions for L2ARC (striped) and SLOG (mirrored).
Some systems, however, where not originally build for ZFS and have
premium HW controller with flash-backed writeback cache and passthrough
capability. I wonder if *on these systems* is better to create an
hardware RAID1 array for OS, L2ARC and SLOG or stick with the MD RAID +
direct ZFS access to L2ARC/SLOG partition approach described above.
Disks comprising the main ZFS pool will *always* be passthrough to
guarantee direct disk access.
PROS: an HW RAID1 array can make use of the writeback cache; it is
simpler to manager (if a disk fails, simply replace it); bootloader will
be automatically mirrored.
CONS: L2ARC will be mirrored rather than striped, auto checksum heal
will be limited/absent (due to ZFS missing direct disk access).
I'm open to suggestions :)
Assyoma S.r.l. - www.assyoma.it
email: g.danti at assyoma.it - info at assyoma.it
GPG public key ID: FF5F32A8
More information about the zfs-discuss