[
https://issues.apache.org/jira/browse/DERBY-4565?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Rick Hillegas updated DERBY-4565:
---------------------------------
Attachment: derby-4565-02-aa-useWaitTimeout.diff
Attaching derby-4565-02-aa-useWaitTimeout.diff. This patch eliminates the retry
count. Instead, we now use the lock wait timeout to bound how long we attempt
to allocate a new cache of sequence numbers. That is, we raise a "too much
contention" exception only if we have looped longer than the timeout set by
derby.locks.waitTimeout.
With this change, the "too much contention" exceptions disappear. Throughput on
the experiment now comes in around 350 tx/sec.
A couple other experiments:
1) I let only one session at a time try to allocate another block of sequence
numbers in its main execution transaction. This reduced throughput to 100
tx/sec.
2) I tried various larger sizes for the number of values pre-allocated in one
gulp. The maximum chunk I tried was 10,000,000 values. This pushed the
throughput up to 480 tx/sec. Most of that boost was achieved by raising the
chunk size to only 100 values.
3) I do see sessions failing to get a write lock immediately, and then falling
back on their execution transactions.
4) I don't see any lock timeout exceptions in the test case itself. That is,
the test does not log any of these exceptions. Knut, is it possible that what
you are seeing are the subtransactions failing to get a write lock immediately
and then falling back on the parent transaction?
I think this patch by itself is incremental improvement:
A) It eliminates the "too much contention" exceptions.
B) It makes it possible to adjust a pre-existing Derby knob in order to tune
the concurrency of sequence generators.
Touches the following files:
---------------
M java/engine/org/apache/derby/impl/services/locks/LockTable.java
M java/engine/org/apache/derby/impl/services/locks/LockSet.java
M java/engine/org/apache/derby/impl/services/locks/ConcurrentLockSet.java
M java/engine/org/apache/derby/impl/services/locks/AbstractPool.java
M java/engine/org/apache/derby/iapi/services/locks/LockFactory.java
Add a new method to the LockFactory so that the factory can report what the
lock wait timeout is.
---------------
M java/engine/org/apache/derby/impl/sql/catalog/SequenceUpdater.java
Use the lock wait timeout to bound the number of times that a session tries to
pre-allocate a chunk of sequence numbers.
> Create concurrency test to stress sequence generators.
> ------------------------------------------------------
>
> Key: DERBY-4565
> URL: https://issues.apache.org/jira/browse/DERBY-4565
> Project: Derby
> Issue Type: Task
> Components: SQL, Test
> Affects Versions: 10.6.0.0
> Reporter: Rick Hillegas
> Assignee: Rick Hillegas
> Attachments: derby-4565-01-ad-firstRev.diff,
> derby-4565-02-aa-useWaitTimeout.diff
>
>
> Create a concurrency test to find bottlenecks and bugs in sequence generators.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
