Mike,
Thanks for your thought on this, see my comments further down.
Mike Matrigali wrote:
Initial writes to the log are to a preallocated log, but once it
is filled it is possible for there to be writes that extend the log
and thus it is not safe to not sync metadata that tracks the
length of the file.
I agree that this is a problem that needs to be looked into if we would
consider to change from using "rws" to "rwd" for the log files. My main
purpose with sending out the performance numbers was to illustrate that
there on some platforms is a potential for performance improvements and
to get feedback on what issues that needed to be looked into in order to
take advantage of this.
Unfortunately this behavior is hardware, OS and JVM specific, and
the exact
meaning of rws and rwd is left vague in the javadoc that I have read.
The javadoc usually says syncs "metadata" but does not explain what
metadata.
I certainly agree that the javadoc is vague. For RandomAcceFile the
javadoc says [1]:
"The "rws" and "rwd" modes work much like the |force(boolean)|
<http://java.sun.com/javase/6/docs/api/java/nio/channels/FileChannel.html#force%28boolean%29>
method of the |FileChannel|
<http://java.sun.com/javase/6/docs/api/java/nio/channels/FileChannel.html>
class, passing arguments of true and false, respectively, except that
they always apply to every I/O operation and are therefore often more
efficient. If the file resides on a local storage device then when an
invocation of a method of this class returns it is guaranteed that all
changes made to the file by that invocation will have been written to
that device. This is useful for ensuring that critical information is
not lost in the event of a system crash."
It is very unclear if the last sentence refer to only files opened with
"rws" mode or also holds true when opened with "rwd" mode.
When I worked on this issue for another db vendor, direct
OS access usually provide 3 rather than 2 options. The 3 options
were:
1) no metadata sync
2) only sync file allocation metadata
3) sync file allocation metadata and other metadata. The problem
is that other metadata includes the last modified time info
which is updated every write to the file.
What do you mean by "most OS"?
Solaris and FreeBSD :-) I have also tried this on Linux 2.6 (Red Hat
4.0?) , but Linux seems to handle "rwd" the same as "rws" (or rather
files where the application requests to be opened with the O_DSYNC flag
is actually opened with just O_SYNC).
The JavaDoc for RandomAccessFile is also indicating that using rws needs
two updates to the disk: "using "rws" requires updates to both the
file's content and its metadata to be written, which generally requires
at least one more low-level I/O operation." [1] Based on this I also
assumed that this would be a penalty on "most OS".
What OS/JVM are your numbers from?
My numbers where from a machine running Solaris 10 x86 and Sun JVM 1.5.
When the sync option on the log was switched from using full file
sync to "rws" mode tests were run which I believe included linux
(probably only a single version of linux - not sure which) and
XP with sun and ibm jvms (probably 1.4.2 as I think that was the latest
JVM at the time), I think apple OS was also tested but I am not sure.
The first implementation simply switched the
to the "rws" mode but left the log file to grow as needed, "rws" mode
was picked because it is impossible to tell if file allocation metadata
is synced as part of "rwd" so in order to guarantee transaction
consistency the safest mode was picked. Tests were run which observed
if we preallocated the log file then I/O to a preallocated file that
did not extend the file only paid 1 I/O per sync. So work was done
to make most log I/O only happen to a preallocated file, but the logging
system was not changed to guarantee all I/O was to a preallocated file.
Thanks for the background for how and why "rws" was selected. Based on
your observations and the unclear semantics for "rwd" and metadata I
agree that this was a good choice. Still, I think paying the extra cost
of having to do two disk operations per log write on "some OSs" is high
and can make Derby perform worse than some of the other open-source
databases on these OSs.
It is probably worth resurrecting the simple I/O test program, to let
people run on their various JVM/OS combinations. As has been noted in
the past the results of such a test can be thrown way off by the
hardware involved. If the hardware/filesystem has had write cache
enabled then none of these syncs can do their job and transactions are
at risk no matter what option is picked.
Sounds like a very good idea to get data for how various JVM/OS
combinations are handling this.
Also it is more common nowadays for higher end hardware to have
battery backed cache to
optimize the sync case, which then provides instantaneous return from
the sync request but provides safe transaction as it guarantees the
write on failure (I've seen this as part of the disk and as part of
the controller). This particular hardware feature works VERY well for
the derby log I/O case as the block being synced for the log file
metadata tends to be the same block over and over again so basically
the cache space for it is on the order of 8k.
And it is very common for lower end hardware to have the disk's write
cache enabled to get similar performance. And most users will be very
happy with this and unaware of the consequences until one day their
favorite database is unable to recover after a power failure....
Regards,
Olav
[1] http://java.sun.com/javase/6/docs/api/java/io/RandomAccessFile.html
Olav Sandstaa wrote:
For writing the transaction log to disk Derby uses a
RandomAccessFile. If it is supported by the JVM, the log files are
opened in "rws" mode making the file system take care of syncing
writes to disk. "rws" mode will ensure that both the data and the file
meta-data is updated for every write to the file. On most operating
system this leads to two write operation to the disk for every write
issued by Derby. This is limiting the throughput of update intensive
applications.
I have run some simple tests where I have changed mode from "rws" to
"rwd" for the Derby log file. When running a small numbers of
concurrent client threads the throughput is almost doubled and the
response time is almost halved. I am enclosing two graphs that show
this when running a given number of concurrent "tpc-b" clients. The
graphs show the throughput when running with "rws" and "rwd" mode
when the
disk's write cache has been enabled and disabled.
This change should also have a positive impact on the Derby startup
time (DERBY-1664) and derbyall. With this change the time for running
derbyall goes down by about 10-15 minutes (approximately 10%) :-)
Is there anyone that is aware of any issues by not updating the file
meta-data for every write? Is there any recovery scenarios where this
can make recovery fail? Derby seems to preallocates the log file
before starting using the file, so I think this should not influence
the ability to fine the last data written to the file after a power
failure.
Any comments?
Thanks,
Olav
------------------------------------------------------------------------
------------------------------------------------------------------------
