Re: separate log and structure from user data device?

[pg_xfs@xxxxxxxxxxxxxxxxxx (Peter Grandi)]

[ ... curious complaints about write performance ... ]

>> /* O_DIRECT for "realtime" */

nathans> You don't need O_DIRECT for realtime these days.

>> assert( (fid = open(pn,
>>   O_WRONLY|O_CREAT|O_DIRECT|O_SYNC|O_LARGEFILE)) != -1 );

nathans> Thats a shocking use of assert. :)

Thats too kind :-).

However, I am puzzled as you comment on 'O_DIRECT' but not on
'O_SYNC'. IIRC XFS performance, and in particular the strategy
to used to allocate large extents, is based on hugely delayed
flushing, and 'O_SYNC' would defeat that, resulting perhaps in
the allocation of many small extents, and this might impact
write speed too (more metadata writes). Is this plausible?

[ ... ]