AIO/DirectIO Integration

[clebert.suconic]

I'm integrating libaio on the Journal, and I had to make few changes on my local box to make it work correctly.

When you are dealing with DirectIO, every write has to be aligned to 512 bytes. To make both AIO and NIO work correctly, I'm adding tail bytes to the end of each record to complete a block of 512 bytes.

Say... you are writing a PrepareRecord

On the regular layout (before this change), you would do this:

1 BYTE - Prepare Record
8 BYTES - Long TXID
1 BYTE Done


After the change I'm making.. you will have

1 Byte - Prepare Record
8 Bytes - Long TXID
4 Bytes - Int fillerSize
1 Byte Done
N Bytes to complete the block (as determined by fillerSize)



With the second layout, both NIO and AIO can work on the same layout.

After we add an intermediate buffer on AIO, we will be able to remove these filler bytes.

[clebert.suconic]

I removed the requirement for the fillerSize, as I have talked to Tim.

Now we aways move to the next position just based on the alignment of the File.


I have committed this work into this Branch:

http://anonsvn.jboss.org/repos/messaging/branches/trunk_tmp_aio/

I still have some work to do, like refactoring on few tests that are dealing with record sizes (what will change when using AIO).

And also, I've found some problem with reclaiming files. The counter doesn't work properly in some situation when the commit is sent to another file. I didn't fully understand the counting, but it looks like there is a problem on the ref count.

[timfox]

And also, I've found some problem with reclaiming files. The counter doesn't work properly in some situation when the commit is sent to another file. I didn't fully understand the counting, but it looks like there is a problem on the ref count.

Can you be more specific?

[clebert.suconic]

Tim Fox wrote:
Can you be more specific?

I just added JournalImplTestUnit::testTransactionOnDifferentFiles, and it is failing.

"Cannot find add info 1."

Note: if you remove the reclaiming call the test will pass.

[clebert.suconic]

This is what is happening:


The transactions are being spread multiple files...

And the commit is alone on the last file...

If there is another record (non related to transactionID=1) that file won't be deleted... but all the others are being deleted.

Later when the file is recovered, Commit(1) won't find a previous add.


Maybe those other files shouldn't be released if the commit can't also be released.


Any ideas Tim?

[timfox]

The test is illegal, since you're adding record 1 in the same tx as you're deleting it.

In messaging you can't do this (send and consume in the same tx).

If you delete something in a tx it must already exist in committed (not transactional) state.

Actually the test is useful though, can you keep it but change it to assert that the exception is thrown as it currently is?

[clebert.suconic]

The test is a simple copy of testMultipleTransactionsDifferentIDs that started to fail when I was using AIO (as it changed the positioning of the records a little bit).

If the test is illegl, a the delete should throw an exception, but the test is failing just because of the position of the commit. In fact I managed to change the test in another way and still fails:

 addTx(1, 1, 2, 3, 4, 5, 6);
 updateTx(1, 1, 3, 5);
 commit(1);
 deleteTx(2, 1, 2, 3, 4, 5, 6);
 commit(2);

 // Just to make sure the commit won't be released. The commit will be on the same file as addTx(3);
 addTx(3, 11, 12);

[timfox]

If you can create a legal test that fails, that's fine.

But it's just confusing to work on illegal tests.

[clebert.suconic]

If you can create a legal test that fails, that's fine.

But it's just confusing to work on illegal tests.

I didn't create the test... just isolated the failure from testMultipleTransactionsDifferentIDs.

[timfox]

What I am saying is come up with a *legal* test that fails, and we'll take it from there....