ntfs directory buffering?

[Mike Ball]

We are using a hardware based replication on a disk
array. When we split the mirror and bring the replicated
filesystem up on the second w2k server the filesystem can
be anywhere from a few minutes to 2 days out of sync with
the primary server. does NTFS do any buffering of
directory writes? is there a way to control this?

 

[Pat [MSFT]]

NTFS does support delayed writes, but generally these are committed within a
few minutes (basically as CPU and disk pressure allows), and the OS also
supports write semantics to forcibly write files (i.e. no OS caching
allowed - SQL, Exchange, etc. do this for data integrity). Keep in mind
that the RAID card probably has its own cache as well. If NTFS sends a
write to disk, the controller may respond that the write did occur even
though it is still on the card (i.e. not committed). IDE cards are
particularly vulnerable as they do not universally enforce the write-through
semantics that SCSI does (i.e. it is legal for them to report a write
successful even though the write only made it to the card).

What is concerning is that the disk data may be missing more than a few
minutes. The mirroring is controlled by the HW, so the OS would not even
know that a mirror replication is occurring. It sounds like the firmware is
not keeping the images in-synch. I would recommend checking w/the vendor to
see if there is an updated driver/firmware available.


Pat