-
1
- #1
JosephKarpinski
Programmer
Hi,
We are looking at implementing CIO on Oracle 10G in our organization and have reviewed the case study
"Improving Database Performance with AIX Concurrent I/O".
When comparing the default behavior of AIX using buffered JFS2 versus CIO,
our tests seem to indicate that buffered JFS2 provides better overall performance.
Some of our observations are:
1. While CIO delivered better disk service time, without buffered JFS2, we did a lot more I/O.
2. My best guess is that this is related to Oracle sequential scans of indexes and tables,
that have only a few data blocks at the LRU end of the SGA buffer pool.
With buffered JFS2, OS file buffers act as a secondary SGA area.
Sequential scans of moderate sized indexes and tables, are resolved in memory,
requiring less physical I/O and delivering faster request times as measured by
less Oracle internal wait time for "db file sequential read" and "db file scattered read".
3. To reduce this negative impact of using CIO, we're looking at testing the buffering of
moderately size sequentially scanned indexes and tables in an Oracle "Keep" database buffer pool.
4. With CIO:
Wait time associated with the wait class "Commit" log file sync event was reduced.
Wait time associated with the wait class "User I/O" events "db file sequential read", "db file scattered read" and "read by other session" increased.
5. The answer appears to be, "it depends", on the application profile.
CIO gives superior disk response time, and for heavy I/O bound applications, it's a good fit.
With CIO turned off, the disk response times were good, averaging between 4 and 5 milliseconds.
With CIO turned on, the disk response times were much better, averaging between 1 and 2 milliseconds.
The disk service time measurements, can be misleading,
in that the disk storage subsystem seems to be bias toward
heavy I/O requests, giving better service time to such requests.
Because CIO generated a lot more I/O requests,
the storage subsystem cacheing and prefetching, provided
better service times.
For applications that get good database read ratio's from the combination of database buffers and operating system file buffers, CIO can induce more disk I/O requests, and thus poorer response times and longer batch runs.
It looks like thread52-1140323 asked similar questions.
Have others seen similar issues when using Oracle on AIX with CIO?
We are looking at implementing CIO on Oracle 10G in our organization and have reviewed the case study
"Improving Database Performance with AIX Concurrent I/O".
When comparing the default behavior of AIX using buffered JFS2 versus CIO,
our tests seem to indicate that buffered JFS2 provides better overall performance.
Some of our observations are:
1. While CIO delivered better disk service time, without buffered JFS2, we did a lot more I/O.
2. My best guess is that this is related to Oracle sequential scans of indexes and tables,
that have only a few data blocks at the LRU end of the SGA buffer pool.
With buffered JFS2, OS file buffers act as a secondary SGA area.
Sequential scans of moderate sized indexes and tables, are resolved in memory,
requiring less physical I/O and delivering faster request times as measured by
less Oracle internal wait time for "db file sequential read" and "db file scattered read".
3. To reduce this negative impact of using CIO, we're looking at testing the buffering of
moderately size sequentially scanned indexes and tables in an Oracle "Keep" database buffer pool.
4. With CIO:
Wait time associated with the wait class "Commit" log file sync event was reduced.
Wait time associated with the wait class "User I/O" events "db file sequential read", "db file scattered read" and "read by other session" increased.
5. The answer appears to be, "it depends", on the application profile.
CIO gives superior disk response time, and for heavy I/O bound applications, it's a good fit.
With CIO turned off, the disk response times were good, averaging between 4 and 5 milliseconds.
With CIO turned on, the disk response times were much better, averaging between 1 and 2 milliseconds.
The disk service time measurements, can be misleading,
in that the disk storage subsystem seems to be bias toward
heavy I/O requests, giving better service time to such requests.
Because CIO generated a lot more I/O requests,
the storage subsystem cacheing and prefetching, provided
better service times.
For applications that get good database read ratio's from the combination of database buffers and operating system file buffers, CIO can induce more disk I/O requests, and thus poorer response times and longer batch runs.
It looks like thread52-1140323 asked similar questions.
Have others seen similar issues when using Oracle on AIX with CIO?