[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

Re: Good journal size on data=journal?

On Jun 17, 2003  03:59 +0200, prox vger org wrote:
> Im wondering what would be a good size on the journal file for a
> data=journal ext3 ? I can't find any reference to a "standard" value for
> other then ordered (8192 blocks).

Default (mke2fs) journal size is the same for all journal types.  The
"ideal" journal size is at least 4x amount of dirty (meta)data in 5s
(journal flush interval).  The 4x is because you need enough space for
1 interval of in-progress journaling, 1 interval of committing, and
1 or 2 intervals of flushing, so that when you wrap around your journal
the space is available for use again.

For Lustre data servers (not journal=data mode), the most dirty data they
normally have is approximately 1/1024 of RAM (number of indirect blocks
as a fraction of data blocks), so 4x that size, 1/256 of RAM is great.
These nodes basically only write data and do very little metadata ops.

For metadata- or small-file-intensive ops, you want about 4*64*5*ops/sec
amount of journal space (4x as above, 5s * ops/sec, ~64 journal blocks/op).

If you are doing sync IO, like with NFS, the amount of outstanding dirty
data also depends on the number of active NFS server threads, but as a
rule "use as large a journal as possible".

Cheers, Andreas
Andreas Dilger

[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]