Allow more concurrency in the lock manager

• Key: DERBY-1704
• Type: Improvement
• Status: Closed
• Resolution: Fixed
• Priority: Minor
• Assignee: Knut Anders Hatlen
• Reporter: Knut Anders Hatlen
• Component/s: Services
• Affects Version/s: 10.2.1.6
• Fix Version/s: 10.3.1.4

Description

I have seen indications of severe monitor contention in SinglePool
(the current lock manager) when multiple threads access a Derby
database concurrently. When a thread wants to lock an object, it needs
to obtain the monitor for both SinglePool and LockSet (both of them
are global synchronization points). This leads to poor scalability.

We should investigate how to allow more concurrency in the lock
manager, and either extend SinglePool or implement a new manager.

Comments

[Knut Anders Hatlen]

Just to see how the performance would be affected if the global
synchronization points were eliminated, I split the hash tables in
SinglePool and LockSet into 16 partitions (that is, 16 hash tables),
and used the hash key to decide which partition an object should be
placed in. There was no global synchronization, only synchronization
on the partition.

I have attached graphs for some performance tests with single-record
selects (the entire database was in the page cache). The graphs show
the results on a machine with a single CPU (1cpu.png), one with two
CPUs (2cpu.png) and one with eight CPUs (8cpu.png). They indicate that
there is no significant effect (neither positive nor negative) on the
single-CPU machine, but with multiple CPUs (and multiple threads) the
performance increases when the global synchronization points have been
eliminated.

(Don't pay too much attention to the actual TPS numbers, since the
three machines didn't have identical processors.)

[Knut Anders Hatlen]

Attaching the patch used to partition the hash tables. It is only
attached for reference, not for inclusion. I have not updated or added
comments, and I have not tried very hard to make the names of
variables and methods meaningful. If someone wants to implement a new
lock manager based on the ideas in this patch, please feel free to do
so. FWIW, derbyall ran cleanly with the patch.

[Anders Morken]

While I don't know exactly what indications you've seen
of the lock contention in the lock manager, we've seen the
exact same thing in our investigations in our "SMP
scalability in Derby" project. (We == me and Per Ottar
Ribe Pahr).

Anyway, we used DTrace with a JDK1.6 beta release to
measure monitor contention in a simple embedded derby
scenario consisting of 8 threads doing simple "SELECT A
FROM FOO WHERE A=?; COMMIT;" loops. FOO is a table
consisting of about 100 000 rows of A, an integer and the
primary key, and a CHAR(96) FOR BIT DATA filler.

The dtrace script measured three things: The classes of
objects whose monitors were most contented - in terms
of number of contentions and the wait time threads spent
waiting for their monitors. In addition, it took a snapshot of
the call stack at the point of contention to figure out which
method the contention occured in.

Here's our measurements, we hope they're useful, if only
to confirm that you're onto something. =)

PS: You want a fixed-point font to read this. Point your web
browser at http://folk.ntnu.no/andersmo/monitor_contention.txt
to get a plain old text file. =)

TOP BLOCKING OBJECTS:

Monitor class Contention count
  org/apache/derby/impl/services/locks/ActiveLock 2
  java/lang/ref/ReferenceQueue$Lock 2
  org/apache/derby/impl/services/cache/CachedItem 2
  org/apache/derby/impl/services/reflect/ReflectLoaderJava2 3
  java/io/PrintStream 5
  org/apache/derby/impl/store/raw/data/StoredPage 10
  sun/misc/Launcher$AppClassLoader 10
  org/apache/derby/impl/sql/GenericPreparedStatement 10
  org/apache/derby/impl/store/raw/xact/XactFactory 11
  org/apache/derby/impl/store/raw/xact/TransactionTable 29
  java/util/Hashtable 34
  org/apache/derby/impl/store/raw/data/AllocationCache 36
  org/apache/derby/impl/services/cache/Clock 257
  org/apache/derby/impl/services/locks/SinglePool 577
  org/apache/derby/impl/services/locks/LockSet 7851

TOP BLOCKING OBJECTS BY WAIT TIME:

Monitor class Wait time (ms)
  java/lang/ref/ReferenceQueue$Lock 2
  org/apache/derby/impl/services/locks/ActiveLock 2
  org/apache/derby/impl/services/cache/CachedItem 6
  org/apache/derby/impl/store/raw/data/StoredPage 11
  org/apache/derby/impl/services/reflect/ReflectLoaderJava2 13
  org/apache/derby/impl/store/raw/data/AllocationCache 110
  org/apache/derby/impl/sql/GenericPreparedStatement 120
  org/apache/derby/impl/store/raw/xact/TransactionTable 129
  java/util/Hashtable 198
  org/apache/derby/impl/store/raw/xact/XactFactory 350
  sun/misc/Launcher$AppClassLoader 660
  java/io/PrintStream 709
  org/apache/derby/impl/services/cache/Clock 1380
  org/apache/derby/impl/services/locks/SinglePool 3137
  org/apache/derby/impl/services/locks/LockSet 16194

TOP BLOCKING METHOD SIGNATURES:

  org/apache/derby/impl/services/locks/LockSet.lockObject(Ljava/lang/Object;Lorg/apache/derby/iapi/services/locks/Lockable;Ljava/l
ang/Object;ILorg/apache/derby/iapi/services/locks/Latch;)Lorg/apache/derby/impl/services/locks/Lock;*
               55

  org/apache/derby/impl/services/cache/Clock.find(Ljava/lang/Object;)Lorg/apache/derby/iapi/services/cache/Cacheable;*
               56

  org/apache/derby/impl/services/locks/LockSet.lockObject(Ljava/lang/Object;Lorg/apache/derby/iapi/services/locks/Lockable;Ljava/l
ang/Object;ILorg/apache/derby/iapi/services/locks/Latch;)Lorg/apache/derby/impl/services/locks/Lock;*
               56

  org/apache/derby/impl/services/cache/Clock.release(Lorg/apache/derby/iapi/services/cache/Cacheable;)V*
               65

  org/apache/derby/impl/services/locks/SinglePool.lockAnObject(Ljava/lang/Object;Ljava/lang/Object;Lorg/apache/derby/iapi/services
/locks/Lockable;Ljava/lang/Object;ILorg/apache/derby/iapi/services/locks/Latch;)Lorg/apache/derby/impl/services/locks/Lock;*
               79

  java/util/Hashtable.get(Ljava/lang/Object;)Ljava/lang/Object;*
               84

  org/apache/derby/impl/services/locks/SinglePool.unlock(Ljava/lang/Object;Ljava/lang/Object;Lorg/apache/derby/iapi/services/locks
/Lockable;Ljava/lang/Object;)I
               85

  org/apache/derby/impl/services/locks/SinglePool.lockObject(Ljava/lang/Object;Ljava/lang/Object;Lorg/apache/derby/iapi/services/l
ocks/Lockable;Ljava/lang/Object;I)Z
              103

  org/apache/derby/impl/services/locks/LockSpace.unlockGroup(Lorg/apache/derby/impl/services/locks/LockSet;Ljava/lang/Object;)V
              110

  org/apache/derby/impl/services/locks/LockSet.unlock(Lorg/apache/derby/iapi/services/locks/Latch;I)V*
              153

  java/util/Hashtable.get(Ljava/lang/Object;)Ljava/lang/Object;*
              221

  org/apache/derby/impl/services/locks/SinglePool.lockAnObject(Ljava/lang/Object;Ljava/lang/Object;Lorg/apache/derby/iapi/services
/locks/Lockable;Ljava/lang/Object;ILorg/apache/derby/iapi/services/locks/Latch;)Lorg/apache/derby/impl/services/locks/Lock;
              260

  org/apache/derby/impl/services/locks/SinglePool.unlatch(Lorg/apache/derby/iapi/services/locks/Latch;)V
              264

  org/apache/derby/impl/services/locks/SinglePool.unlock(Ljava/lang/Object;Ljava/lang/Object;Lorg/apache/derby/iapi/services/locks
/Lockable;Ljava/lang/Object;)I
              330

  org/apache/derby/impl/services/locks/SinglePool.latchObject(Ljava/lang/Object;Lorg/apache/derby/iapi/services/locks/Lockable;Lja
va/lang/Object;I)Z
              459

  org/apache/derby/impl/services/locks/LockSet.lockObject(Ljava/lang/Object;Lorg/apache/derby/iapi/services/locks/Lockable;Ljava/l
ang/Object;ILorg/apache/derby/iapi/services/locks/Latch;)Lorg/apache/derby/impl/services/locks/Lock;*
              559

  org/apache/derby/impl/services/locks/LockSet.unlock(Lorg/apache/derby/iapi/services/locks/Latch;I)V*
             1714

  org/apache/derby/impl/services/locks/LockSet.lockObject(Ljava/lang/Object;Lorg/apache/derby/iapi/services/locks/Lockable;Ljava/l
ang/Object;ILorg/apache/derby/iapi/services/locks/Latch;)Lorg/apache/derby/impl/services/locks/Lock;*
             2840

[Anders Morken]

Oh, forgot to mention that the tests were run on a SMP machine - an old Sun Enterprise 450 with 4x400MHz CPUs (and a bit error on bit 12 on memory module 1701 - yay ECC. ;-). The Derby release used was 10.1.3.1 - (417277), the official release from the web page. If time permits we'll see about testing Knut Anders' patch with DTrace as well.

[Knut Anders Hatlen]

I am planning to partition the hash table in LockSet along the lines
of the split-hashtables.diff patch.

For the other hash table on which there is monitor contention (the one
in SinglePool), I think it is better to remove the hash table
entirely. That hash table maps a compatibility space to a
LockSpace. Since there is a one-to-one mapping between compatibility
spaces and LockSpaces, we could change it so that each compatibility
space is a LockSpace object and remove the need for the hash
table. This would (a) remove the contention problem, and (b) save some
CPU cycles because we don't have to go through the hash table.

If this sounds OK, I will create two sub-tasks: one for partitioning
the table in LockSet, and one for removing the table in SinglePool.

[Daniel John Debrunner]

I haven't had time to look at the patch but the original intention was for there to be a multi-thread lock manager
using multiple LockSets, e.g. a MultiPool implementation instead of a SinglePool implementation. Just FYI.

With the multiple synchronization how is deadlock detection to be handled?

[Knut Anders Hatlen]

Thanks. Do you think it would be better to leave SinglePool as it is
and create a new (perhaps optional) MultiPool implementation instead?


I haven't looked very closely at the deadlock detection code yet, but
I hope it won't be necessary with too many changes. One possibility is
to obtain the synchronization locks on all partitions before the
waiters graph is built. Of course, some precautions are needed in
order to avoid java-level deadlocks when are multiple synchronization
locks possibly obtained in different order.

--
Knut Anders

[Knut Anders Hatlen]

In the patch I used when running the tests, monitor contention on LockSet was reduced by having multiple hash tables in each LockSet. I think it might be cleaner to have multiple LockSets, where each LockSet contains one hash table. Will look more into it and report back. In the meantime, I plan to submit some cleanup patches that hopefully will make the final patch smaller and easier to understand regardless of which approach is chosen.

[Knut Anders Hatlen]

Attached cleanup1.diff.

I observed that LockSet extends Hashtable, but all calls to Hashtable's methods are already synchronized on the LockSet, so it could just as well use a HashMap. The attached patch makes LockSet contain a HashMap instead of extending Hashtable. I ran some tests with the DERBY-1961 test client (single-record select on a dual-cpu machine, Solaris 10, Java SE 6) and saw ~2% throughput improvement for a single client and from 15% to 25% improvement for multiple (2-64) clients.

Derbyall and the JUnit tests passed on Solaris 10/Java SE 6. Reviews would be appreciated.

[Dyre Tjeldvoll]

cleanup1.diff looks good, but I have a couple of questions:

LockSet has was declared "public final", but is now just "final". Is it intentional?

In general, I have a hard time understanding the multithreading-related comments used in Derby (not specific to this patch). For example, the comment for LockSet says:
"MT - Mutable - Container Object : Thread Safe". I really don't understand what that means. Is it still true when LockSet no longer inherits from a synchronized container?

I think what is meant here is that only those methods declared "synchronized" are MT-safe. A client wanting to use any of the other methods must synchronize on the LockSet object. Correct?

The method addWaiters(Dictionary) is not synchronized, and the comment says "MT - must be synchronized on this LockSet object". But in Deadock.getWaiters(LockSet) it is called without synchronization. Presumably this is ok since the old code iterated using an enumeration over an unlocked LockSet, (unless the lock is higher up the stack, and cannot be seen in the diff). But a comment would have been nice :)

[Knut Anders Hatlen]

Thank you for looking at the patch! I'll try to answer your questions below.

1) The change of LockSet from "public final" to "final" was intentional. LockSet is an internal implementation class that is not supposed to be accessed from other packages.

2) I'm not sure what the exact meaning of the MT comment is. I would assume that it meant something like "calls to public (or non-private) methods of this class can be regarded as atomic operations". I'm not sure that this statement is completely true before the patch, but it is definitely not true after the patch, so the comment should be updated.

3) getWaiters() is a private helper method for Deadlock.look() which is only invoked from inside a synchronized block in LockSet, so getWaiters() is in fact always synchronized on the LockSet. This would probably have been clearer if Deadlock.look() had one of those infamous MT comments. :) I'll add one.

[Knut Anders Hatlen]

I have attached v2 of cleanup1. I went through LockSet's methods to determine which of them were not thread safe without extra synchronization and updated their javadoc comments. Changes from the first version:

* MT comment for LockSet class changed to "MT - Mutable - Container Object : All non-private methods of this class are thread safe unless otherwise stated by their javadoc comments."

* Added javadoc comments to LockSet.oneMoreWaiter() and LockSet.oneLessWaiter() with information about required synchronization.

* Removed what seems to be a false comment in LockSet.anyoneBlocked() ("no synchronization needed because reads of ints are atomic"). It is true that reads of ints are atomic, but synchronization is still needed to avoid that an old cached value is read instead of the real value.

* Declared LockSet.anyoneBlocked() as synchronized to avoid the problem mentioned above. Currently, I don't think it is a problem since all calls to anyoneBlocked() come (indirectly) from the unit tests T_AccessFactory and T_LockFactory, but it would be good to fix it anyway.

* Added javadoc and MT comment to Deadlock.look().

I have started a new run of derbyall.

[Dyre Tjeldvoll]

New version looks good. +1

[Knut Anders Hatlen]

Thanks for the review! Committed cleanup1.v2 with revision 498999.

[Knut Anders Hatlen]

Committed cleanup2.diff with revision 499316. The patch removes an unused variable, an unused import and an unused method.

[Knut Anders Hatlen]

LockSpace extends Hashtable, but no other classes use its Hashtable methods (except one call to isEmpty()), so the Hashtable could just as well be a private member. Since all the non-private methods of LockSpace are synchronized, the Hashtable could be replaced with an unsynchronized HashMap. cleanup3.diff makes these changes. Derbyall and JUnit passed.

[Knut Anders Hatlen]

The results shown in the attached graphs are not valid for trunk after the commit of the DERBY-2107 patch. The single-record select load that was used is more latch-intensive than lock-intensive, and when the latches no longer use the lock manager, the effect of splitting the hash tables in the lock manager is small, even with multiple CPUs. There still is a positive effect for more lock-intensive loads though (for instance scans).

Since there are two global synchronization points in the lock manager, I'll file subtasks so they can be discussed separately.

[Anders Morken]

In slightly related work, we've recently done a little testing with a single-record select load on a trunk patched with DERBY-2107 as well as a port of the patches to split the hash tables in the lock subsystem that you included here.

With a separate table and index for each thread (to remove latch contention and lock waits from the equation) we got an overall throughput increase of about 16% compared to a "single table for all threads" run (given a cache large enough to maintain the database in-memory), and found that org.apache.derby.impl.services.cache.Clock.find()/release() caused about 5 times more contention than the synchronization in LockSet.lockObject() and LockSet.unlock(). That might be an indicator of where to apply the next push - and validates the "split hashtable" approach for this workload. =)

For our part, we're proceeding to hack at latching. =)

[Knut Anders Hatlen]

Committed cleanup3.diff with revision 507428.

[Knut Anders Hatlen]

Attaching a tiny cleanup patch (cleanup4). LockControl.popFrontWaiter() is identical to removeWaiter(), only that it calls remove() with the constant 0 instead of using a parameter. The patch makes popFrontWaiter() forward calls to removeWaiter(). The tests passed.

[Knut Anders Hatlen]

Committed cleanup4.diff with 517588.

[Knut Anders Hatlen]

cleanup5.diff contains these changes:

ActiveLock.java: removed some unused imports
Lock.java: made the class package private
LockControl.java:
  - made the class final and package private
  - addLock() had an if which tested (!grantLock && ...) at a place where grantLock always would be false. Removed the !grantLock part of the condition.
  - addWaiter(), removeWaiter() and popFrontWaiter() took a List argument called waiting. Since the argument always was identical to a private member with the same name, I removed "List waiting" from the parameter list.

Derbyall and the JUnit tests ran successfully.

[Knut Anders Hatlen]

Attaching a new small cleanup patch (cleanup6) for LockSet.java. LockSet has a private Hashtable (lockTraces) into which it puts a Throwable each time a thread must wait for a lock (only if deadlock tracing is enabled). The values in the hash table are however never used, so the hash table and the code that updates it could be removed. Derbyall and the JUnit tests ran cleanly with the patch.

[Knut Anders Hatlen]

Committed cleanup5 with revision 518052.
Committed cleanup6 with revision 518073.

[Knut Anders Hatlen]

All sub-tasks completed. Closing the issue.