Hi Mike,
the issue I'm having is I can't find a way to tell the network server
what StorageFactory to use from the network client driver.
On the embedded side of things, I just define a new subsubprotocol when
I start java with:
-Dderby.subSubProtocol.memory=org.apache.derby.impl.io.MemoryStorageFactory
(alternatively I can change some of the engine code and register it as a
persistent service which results in the same issue)
then tell it to use that StorageFactory when I connect to the database, ie:
jdbc:derby:memory:dbname;attributes
However, there does not appear to be an option to explicitly tell the
network server which storagefactory to use.
In the section entitled "Accessing the Network Server by using the
network client driver" at
http://db.apache.org/derby/docs/dev/adminguide/adminguide-single.html,
it looks like the driver syntax does not support the above form of
connection URL. There does not appear to be an attribute for specifying
the storagefactory either.
I am looking for confirmation that this is the case.
If it is, I think the easiest way to add this functionality is another
connection attribute. This should avoid any code changes to the drivers.
-------
As for my performance issues, I can tell the disk to turn off the write
cache, which it promptly ignores :) (Sound familiar?
http://weblogs.java.net/blog/davidvc/archive/2005/10/the_story_of_th.html)
Good suggestions to try durability=test mode and make the page cache
gigantic. I will test them versus my in-memory implementation.
Thanks :)
I did figure out what was giving me the unexpectedly poor performance
increase while running in-memory. A little error on my part. I'm now
getting 2-3x better performance than Dir4StorageFactory :). Though this
number is entirely dependant on RAM speed, OS and disk speed so it will
vary between machines. My testbed was Gentoo Linux, kernel 2.6.13, JDK
1.4.2.08, 7200 RPM ATA hard drive w/8MB cache, 768MB of DDR266 RAM and
an AMD 2400+ CPU.
Thanks for your advice,
Stephen Fitch
Acadia University
Mike Matrigali wrote:
I am not sure I understand your question about network client/server and
Storage factory. The Network layer is way above the StorageFactory
layer, so any change you make that makes an embedded client run in
memory will automatically also be used if you access it using the
network server interfaces (the in memory portion will always be on the
server side of the client/server interface).
Your numbers don't surprise me if your disk is not really supporting
synchronous I/O for log and data. It may also be interesting to compare
your implementation numbers with running the existing server under
durability=test mode, especially if you ever figure out how to get your
disk to allow you to sync the log file.
The areas that are going to cause Derby to go slow:
o sync of log file at commit time (write cache enabled disk disables
this sync)
o sync of data files at checkpoint time (write cache disables this sync)
o sync of data files when db cache gets full (write cache disables, for
fair comparison you should make cache as big as database if you want
to compare with your in memory implementation)
o read of uncached data from disk into cache - I guess this never
happens in your implementation as all data has to come about from updates.
Stephen Fitch wrote:
Hi Norbert,
I have it working in the trunk embedded client in gentoo linux under
jdk 1.4.2.08 but it should theoretically run under jdk 1.3.1 and any
recent release of derby. However, as far as I can tell the network
client/server doesn't support alternate implementations of
StorageFactory. I'm still trying to track down if this is the case.
If it is, another connection URL attribute may have to be added for
in-memory to work on the network side of things.
I'm having issues trying to test the performance as well since I'm
competing with derby's page cache, java.nio's and my hard drive write
cache (which I can't seem to turn off).
*VERY* preliminary numbers indicate a 10-20% performance increase on
inserts which is well below my expectations. I'm going to code up
some tests for other operations today though.
If anyone's interested in trying the code out or going over it to look
for areas of improvement it would be much appreciated.
I suppose the best way to distro it would to email a svn diff patch
and source code for the new classes to the mailing list? I'm not
looking to get it added to svn though, as it still needs work.
Regards,
Stephen Fitch
Acadia University
Toth-Gati Norbert wrote:
Hi Stephen,
So now this means the support for in-memory storage is completed.
I will give it a try. But glad you finally got to the end. Good job!
Regards,
Norbert
----- Original Message ----- From: "Stephen Fitch (JIRA)"
<[email protected]>
To: <[email protected]>
Sent: Tuesday, October 25, 2005 8:33 PM
Subject: [jira] Created: (DERBY-646) In-memory backend storage support
In-memory backend storage support
---------------------------------
Key: DERBY-646
URL: http://issues.apache.org/jira/browse/DERBY-646
Project: Derby
Type: New Feature
Components: Store
Environment: All
Reporter: Stephen Fitch
To allow creation and modification of databases in-memory without
requiring disk access or space to store the database.
--
This message is automatically generated by JIRA.
-
If you think it was sent incorrectly contact one of the administrators:
http://issues.apache.org/jira/secure/Administrators.jspa
-
For more information on JIRA, see:
http://www.atlassian.com/software/jira
