Oracle gives Log read is SYNCHRONOUS though disk_asynch_io is enabled!

Question

Hi,
We have a Oracle 11.2.0.2.0 64 bit database on a SuSe 10 server (with NFS mounts) and in some trace file I did found the following message:
Log read is SYNCHRONOUS though disk_asynch_io is enabled!
I'm puzzled because Async Io is enabled:
filesystemio_options is set to SETALL and disk_async_io is set to TRUE. When I query:select file_no,filetype_name,asynch_io from v$iostat_file;
I get the following:
FILE_NO FILETYPE_NAME ASYNCH_IO
0 Other ASYNC_OFF
0 Control File ASYNC_OFF
0 Log File ASYNC_OFF
0 Archive Log ASYNC_OFF
0 Data File Backup ASYNC_OFF
0 Data File Incremental Backup ASYNC_OFF
0 Archive Log Backup ASYNC_OFF
0 Data File Copy ASYNC_OFF
0 Flashback Log ASYNC_OFF
0 Data Pump Dump File ASYNC_OFF
all datafiles: async_on!
The NFS mount options for all filesystems are correct: rw,v3,rsize=32768,wsize=32768,hard,nolock,timeo=600,proto=tcp,sec=sys
Is IO for log file allways synchrounous? The reason for my question is that we see from time to time the following Awr information:
Top 5 Timed Foreground Events
Event Waits     Time(s)     Avg wait (ms)     % DB time     Wait Class
log file sync     62,576     32,908     526     67.98     Commit
db file sequential read     539,144     11,033     20     22.79     User I/O
DB CPU          3,800          7.85     
enq: TX - row lock contention     375     693     1848     1.43     Application
Log archive I/O     667     59     89     0.12     System I/O
During this high log file sync users complain about bad performance.
Any ideas?
regards,
Ivan

Answer 1

>
I think this is correct. The concept behind commit processing, where your session must hang following commit until the logfile write is complete, would seem to require synchronous writes. If the write were asynchronous, in effect Oracle could return a "commit complete" message before the redo was written to disc, so it would then be theoretically possible to lose a committed transaction.
>
Not necessarily. Though commit's logical write must be "synchronous" or "assured", it is not necessarily that it must use synchronous i/o calls for it.
For example Log Writer can issue 3 asynchronous writes to 3 logfiles simultaneously and then wait (or do other stuff like next buffer population) until all 3 finish, then if all 3 done successfully return control to user process.
From a user process perspective the commit still look solid and "synchronous". But here Log Writer benefits from simultaneous parallel write to 3 disks. Plus Log Writer meantime can do some other activity with buffers or whatever. Other words it is 3 times faster versus it do that in 3 sequential synchronous i/o calls.
to prove my point I run strace on lgwr process and run some quite big transactions.
[root@oracledev ~]# ps -ef|grep lgwr
# root       525 32728  0 13:41 pts/2    00:00:00 grep lgwr
oraerd    7436     1  0 Jan27 ?        00:08:26 ora_lgwr_ERD
# strace -p 7436
io_submit(20971520, 2, {{...}, {...}})  = 2
io_getevents(20971520, 2, 1024, {{...}, {...}}, {...}) = 2
times(NULL)                             = 749695972
times(NULL)                             = 749695972
times(NULL)                             = 749695972
semctl(294917, 96, SETVAL, 0xfff0a600)  = 0
times(NULL)                             = 749695972
times(NULL)                             = 749695972
times(NULL)                             = 749695972
io_submit(20971520, 2, {{...}, {...}})  = 2
io_getevents(20971520, 2, 1024, {{...}, {...}}, {...}) = 2
times(NULL)                             = 749695973
times(NULL)                             = 749695973
times(NULL)                             = 749695973
semctl(294917, 88, SETVAL, 0xfff0a600)  = 0
times(NULL)                             = 749695973
times(NULL)                             = 749695973
semtimedop(294917, 0xfff0c8e0, 1, {...}) = 0
...Note io_submit() and io_getevents() calls - it is Async I/O.
I did not see other I/O like write().
Oracle 10.2.0.5 on Linux on IBM POWER
uanme -a
Linux oracledev 2.6.18-308.1.1.el5 #1 SMP Fri Feb 17 16:51:00 EST 2012 ppc64 ppc64 ppc64 GNU/Linux
Edited by: Mark Malakanov (user11181920) on Mar 5, 2013 1:41 PM