DBWR performance

Did anybody ever push DBWR to reach max disk write speed?

I have /dev/md0, /dev/md1 and /dev/md2 with identical configuration:
raid1 on 2 3TB hdd, 6 hdd total. I can reach 180MB/sec write speed on
each of them running 3 bonnie++ concurrently. Filesystem is xfs on all
of them.

Data goes on /dev/md0, redo on /dev/md1 and undo on /dev/md2.

Execute insert /*+ append */ and write speed on /dev/md0 goes to
150-160MB/sec as seen by iostat. Good, this is close. Obviously
nothing goes to /dev/md1 and /dev/md2.

Execute insert (no append) and dbwr write speed on /dev/md0 is 0MB/sec
and lgwr on /dev/md1 reaches 130MB/sec until db block buffers are
full. Then dbwr speed is only 50MB/sec and it also holding back lgwr
at 50MB/sec.

To me it looks like dbwr is incapable of writing more then 50MB/sec
and holding everything back.

I would expect write speed on /dev/md0 and /dev/md1 to be 180/180 MB/
sec, not 50/50 MB/sec.

Adding >1 db writers or >1 slaves does not change this picture. Async
io is on, direct io is on. DB block buffers 4GB. Test table size
~10GB.

Config: fedora 16 64bit, oracle 11.2, xfs, 32GB ram, AMD FX-8120.

Andrew
.

Relevant Pages

  • Re: DBWR performance
    ... It does not look like LGWR is an issue, ... I increased db block buffers to 24GB - much bigger ... and dbwr writes nothing to hdd during normal insert, ... the redo protecting that block. ...
    (comp.databases.oracle.server)
  • Re: DBWR performance
    ... | It does not look like LGWR is an issue, ... | then DBWR. ... if it's a normal insert you will also be generating undo, ... Jonathan Lewis ...
    (comp.databases.oracle.server)
  • Re: DBWR performance
    ... Then dbwr speed is only 50MB/sec and it also holding back lgwr ... the redo protecting that block. ... effect by populating the table with rows that are 80 bytes long, ...
    (comp.databases.oracle.server)
  • Re: DBWR performance
    ... Write speeds: ... lgwr - 130MB/sec ... dbwr - 150MB/sec ... Still no cigar :-). ...
    (comp.databases.oracle.server)