Free buffer waits

Question

We are seeing some free buffer waits contention and I wanted to get some input from the forum participants if you have time to address it.
It is on an HP-UX 11.31 Itanium system running 11.2.0.3 of Oracle. This is a datawarehouse staging database and there are about 60 merge statements running in parallel (themselves also doing parallel dml) doing several million updates against 60 different tables with a total rowcount around 2 billion.
The dev team is putting in a more efficient update statement so that may in itself resolve these issues but still I'd like to see if the number of db writer processes we have makes sense. As far as I know we do not have asych io configured on our OS due to some bugs we have seen in the past. The server has 14 cpus and 95 gig of memory. Here are the top 5 events from our AWR report:
Top 5 Timed Foreground Events
Event                   Waits      Time(s)  Avg wait (ms) % DB time Wait Class
free buffer waits       319,324    261,188            818     46.08 Configuration
db file parallel read   134,710    62,404             463     11.01 User I/O
DB CPU                             60,818                     10.73  
db file sequential read 11,783,603 26,032               2      4.59 User I/O
write complete waits    4,015      13,828            3444      2.44 Configuration Does it make sense that I should increase the number of db writers?
db_writer_processes=4 in our system. With 14 cpus this is supposed to be enough as I understand it. But we have 60 dedicated server processes doing updates and only four db writer processes doing writes so it kind of makes sense to increase the number of writers.
I'm researching this on my own but I would appreciate any input you have on this issue.
Thanks,
Bobby

Answer 1

Bobby Durrett wrote:
Well, I think I'm back to increasing the number of db writers. Here is part of Oracle note 139272.1: HP-UX: Asynchronous i/o
It may help - but I'm doubtful.
If nothing else, the comments in the note look as if they're at least 10 years old. (Note, in particular, that one of the references at the end of the note refers to a document about Oracle 8.1.7; I'm also slightly surprised by the lack of reference to I/O Slaves which were introduced to emulate async I/O for platforms that didn't support it; (update) and then there's a reference to an HP-UX patch is for 11.00, and descriptions of what to do for Oracle versions 7.3 and earlier!)
This serial operation can lead to a i/o bottleneck. There are two ways to counteract this:
a. configure multiple DBWR processes
b. use asynchronous i/oYou haven't told us anything about the db file parallel write times yet. If they're slow then increasing the number of db writers is likely to make them slower.
On the information you've given so far, I'd investigate setting filesystem_io_options to setall, in combination with mounting the file system with convosync={whatever it was}, as I suspect locking at the file level within the filesystem cache. (I'm not sure whether you need to set mincache=direct as well to see the write benefit - the problem is that your 11M single block reads might slow down because they would be bypassing the filesystem cache.)
It's at this point on a client site that I tend to get into detailed discussion with the System Admin to understand exactly what filesystem and O/S options are currently available and exactly what they do. These things are too variable (and often badly understood) between operating systems and versions.
Regards
Jonathan Lewis
Edited by: Jonathan Lewis on Apr 24, 2013 9:38 AM