Ehcache BigMemory: Simple High Availability, Even Simpler

My collegue Jason @ Terracotta did a nice post on using Ehcache with JRuby; which led to a discussion on a long list of features we implemented for Enterprise Ehcache (check out the discussion thread on Jason's Blog).

Adding to Jason's feature list, I would like to discuss HA (High Availability) in Ehcache and explain why our BigMemory product makes tuning HA even simpler. Lets review how we do HA in Enterprise Ehcache. In Enterprise Ehcache, clients going down is no big deal since the data is present on the servers as well. Our HA focuses on protecting our servers.  In Enterprise Ehcache you can define one or many server groups. Each group consists of a cluster of servers. The cluster has to decide which node is going to be the active server. This is decided by having an election where a node is selected to be the active server. The rest of the nodes in the cluster are waiting in passive standby ready to take over if the active node fails.

In order to actually detect when a failover needs to happen, we wrote a configurable HC (Health checker).  Our HC detects errors that won't show up as a normal network disconnect or failure, such as a network cable being pulled. Because Enterprise Ehcache is written in Java, we also had to deal with long GCs. So we designed our HC to detect long GCs as well.

Based on your use case, you may want to change the HC settings depending on what your tolerance is for network disruption and long GCs. Before you go about changing your settings, you might want to check out these files:

$TERRACOTTA_KIT/platform/config-samples/tc-config-healthchecker-aggressive.xml
$TERRACOTTA_KIT/platform/config-samples/tc-config-healthchecker-aggressive.xmltc-config-healthchecker-development.xml
$TERRACOTTA_KIT/platform/config-samples/tc-config-healthchecker-aggressive.xmltc-config-healthchecker-production.xml


Depending on what you're doing and what your requirements are, picking one of the settings above should be suffice.

Now let's discuss these properties:

l2.healthcheck.l2.ping.idletime=3000
l2.healthcheck.l2.ping.interval=1000
l2.healthcheck.l2.ping.probes=2
l2.healthcheck.l2.socketConnectTimeout=5
l2.healthcheck.l2.socketConnectCount=2


Above are the properties you have to work with for HC.  The HC starts off using the ping.idletime. This is the maximum amount of time that can elapse between the last time data was received from the corresponding node. In this case the idletime is 3000 milliseconds, after which the HC notes "Hey, didn't receive any data from the corresponding node, I should check on that node."

To check on the health of the node, it tries to ping the node in intervals, in which the interval length is defined by ping.interval. You can push this number down to get more granularity. If the corresponding node doesn't respond within the ping.interval then it either tries to probe again because the ping.probes countdown hasn't completed, or it checks socketConnectCount and see if its allowed to make any more socket connections. If not, it declares the corresponding node DEAD.

In the example above, since the socketConnectCount is set to 2, it will try to make another socket connection. If it cannot make the socket connection within (socketConnectTimeout * pingInterval) ms, then it will declare the node DEAD. In our example, the interval length 5000 ms. Once it established a connection it will repeat the ping probe cycle again.

The maximum time it will take the HC to detect a network disruption is ( ping.idletime + (ping.probes * ping.interval) + (socketConnectTimeout * ping.interval) ) ms. If the problem is longGC, then the connection will happen, but the pings won't receive a response. The maximum time HC takes to detect a long GC is ( socketConnectCount * ( ping.idletime + (ping.probes * ping.interval) + (socketConnectTimeout * ping.interval) ) ) ms.

If you have a short tolerance for network disruption, but your ok with having lengthy long GCs, then you can decrease the ping.idletime and increase the socketConnectCount; you tune based on your tolerances. Here's some detailed documentation on the HA settings.

With BigMemory in our server FORGET ALL THAT IS WRITTEN ABOVE.

Our HC has all these different properties because we had to be tolerant of Long GCs. When a node is in long GC, it will make a socket connection but not be able to complete the ping probe cycle. but now with BigMemory you probably don't ever need to change these settings, unless you have people tripping over your network cables.

Unlike Long GCs, network disruptions is something you probably know about and its easier to guess what that tolerance should be. Not having to tune for long GCs makes HC configuration simple. You only need to tune for YOUR own environment (i.e. crappy network, or clumsy workers) and not for something that is specific to Java (Long GCs).

Imagine what it can do for you. Check out our beta here.

BigMemory: Followup Q and A

I got quite a few responses to my post on BigMemory in the Terracotta Server. It seems like people are quite confused on what it actually is.

Here's some answers to a few questions I received:

1. Why can't they (Terracotta) put garbage collector on another cpu core and gain performance?

I think there is a misunderstanding about the cost of Garbage Collection. The Full GC pause (which is when all application threads are paused) is what the GC problem in Java is all about. It is tolerable when your Heap is 1-2 GB. But anything beyond that you get 4,5,8 seconds GC pauses. Besides, if you don't run ParallelGC then it will use one core anyway. But you DO want to have your garbage collector using all the cores so it will complete faster and have less pauses.

2. (In References to the question above) Then put it on another thread and how about pausing one thread at a time ?

Again this is not possible AFAIK to do with the Sun/Oracle JVM. Also, Full GC Pauses are a necessary evil for the GC algorithm they are using. Even if this was possible, it would not solve the problem of unpredictability.

3. I can't believe there are no GC pauses ... or you guys might have made memory management solution like an OS in java.

The idea of have direct memory allocation in Java is no big secret. There is an -XX:MaxDirectMemorySize flag to tell the JVM how much direct memory to allocate.  The value add of Terracotta is to use this direct memory space in a way that is fast and does got fragment.

4. Using direct memory allocated by the JVM is useless because it is so much slower then the Heap.

Access to direct memory is NOT slower than Heap. There are two things that contribute to the perceived slowness of direct memory. Serializing and deserializing data to and from direct memory; and allocating and cleaning up direct memory buffers. At Terracotta we solved the direct memory and cleanup problem. On the Terracotta Server we don't pay for the serialization/deserialization cost. On Enterprise Ehcache (unclustered) we do pay a serialization/deserialization cost, but compare this CPU cost to having to deal with Full GC Pauses on the Heap. The tradeoff is well worth it. Besides BigMemory using the Heap as part tier storage strategy; Heap to OffHeap to Disk. It's an age old principle in computer science (think Virtual Memory). We avoid the serialization/deserialization cost for frequently used objects by having those in Heap, then having a big part of your cache on OffHeap to avoid long FullGC and the rest spilling over to disk.

For the additional CPU cost what you get in return is predictable latency and speed with all the memory your Java process desires. Find an app where you do see Full GC pauses and checkout the beta to see for yourself.

Performance Tuning Hibernate Second-level Cache

Lately I've been spending ALOT of time performance testing and tuning of a hibernate app using a Terracotta implementation of Hibernate Second-level Cache. I thought it would be interesting to do a series of posts that walk through different use-cases and how we performance tuned them. While doing so, I thought it would also interesting to demonstrate various tools that can help with the tuning.

Before we continue, I think it is worth while to read my colleagues posts on performance testing here, here, and here. Their advice seems simple enough. But trust me, it's really easy to fall off the wagon when it comes to this stuff. The work itself is very tedious, so it's easy to convince yourself that "Yeah, this patch is it, this is the final run", or "I know the bottleneck is the CPU for sure! Why bother looking at the memory and/or disk characteristics." Assumptions and unfounded optimism can easily get you rat-holed to a dead end. Bottom line: don't assume anything and collect everything.

The project that we are going to look at is a simple hibernate app. The app itself just warms up the 2nd-level cache and then accesses the cache, and depending on the use case writes, in a non-partitioned fashion in the following modes: read-only, read-mostly and read-write.

the various tools I used to tune is:
- nmon a basic tool to look at cpu, memory, disk and network metrics.
- jmap,jhat,jstat to analyze heap and garbage collector metrics, these tools are included in the jdk.
- VisualVM, very cool tool to look at JMX Beans, nmon type characteristics as well as a light-weight thread monitoring tool.
- Terracotta Dev Console, great tool for looking at terracotta specific characteristics as well as general info (cpu, memory etc.)
- Good old fashion thread dumps (kill -3 java_pid)

Ok, so that's our test and tools. More of the nitty-gritty to come :)

Terracotta Cluster Events: A simple example

My friend instant messaging me this weekend about this app he is writing using Terracotta. He is concerned about when a node leaves/joins and terracotta cluster and was doing some really weird stuff with the JVM Shutdown hook. I decided to do a quick blog post about Terracotta Clustered Events which is a simple way for an app to listen for terracotta cluster wide events.

So first i'll generate a new project using the pojo-archetype:


mvn org.apache.maven.plugins:maven-archetype-plugin:2.0-alpha-5:generate \
-DarchetypeGroupId=org.terracotta.maven.archetypes \
-DarchetypeArtifactId=pojo-archetype \
-DarchetypeVersion=1.5.0-SNAPSHOT \
-DgroupId=org.terracotta.examples \
-DartifactId=clusteredapi-examples \
-Dversion=1.0.0 \
-DremoteRepositories=http://www.terracotta.org/download/reflector/maven2


The the following prompt appears:


[INFO] Scanning for projects...
[INFO] ------------------------------------------------------------------------
[INFO] Building Maven Default Project
[INFO] task-segment: [org.apache.maven.plugins:maven-archetype-plugin:2.0-alpha-5-parent-patch-SNAPSHOT:generate] (aggregator-style)
[INFO] ------------------------------------------------------------------------
[INFO] Preparing archetype:generate
[INFO] No goals needed for project - skipping
[INFO] Setting property: classpath.resource.loader.class => 'org.codehaus.plexus.velocity.ContextClassLoaderResourceLoader'.
[INFO] Setting property: velocimacro.messages.on => 'false'.
[INFO] Setting property: resource.loader => 'classpath'.
[INFO] Setting property: resource.manager.logwhenfound => 'false'.
[INFO] [archetype:generate]
[INFO] Generating project in Interactive mode
[INFO] Archetype repository missing. Using the one from [org.terracotta.maven.archetypes:pojo-archetype:1.5.0-SNAPSHOT] found in catalog local
[INFO] Using property: groupId = org.terracotta.examples
[INFO] Using property: artifactId = clusterapi-example
[INFO] Using property: version = 1.0.0
[INFO] Using property: package = org.terracotta.examples
Define value for property 'description': :


Now enter in a description, I typed in "This is an example project to demonstrate java clustered api" and then was prompted again and Typed in "Y"
Now we have a project to test out terracotta clustered api.

The clustered event api interfaces are available from the terracotta api jar, let's include that in our pom.xml


<dependency>
<groupid> org.terracotta.api</groupid>
<artifactid>api</artifactid>
<version>1.0.0</version>
</dependency>

Change the process class to have a count of 4, so we run 4 nodes from our tests.


<plugin>
...
<processes>
<process nodename="app" count="4" jvmargs="-Xmx20m">
<classname>org.terracotta.examples.App
</classname>
</process>
</processes>
</plugin>



Now that we have our app I want it to be clustered aware. In this example, I'm not going to do anything interesting except print out when node's joins and leaves. So I added the following listener and implementation to my existing App class:



public class App implements DsoClusterListener {

@InjectedDsoInstance
private DsoCluster cluster;

public void registerListener() {
this.cluster.addClusterListener(this);

}

...
public void nodeJoined(DsoClusterEvent dsoclusterevent) {
System.out.println("nodeJoined Event about node: " + dsoclusterevent.getNode());
}

public void nodeLeft(DsoClusterEvent dsoclusterevent) {
System.out.println("nodeLeft Event about node: " + dsoclusterevent.getNode());
}

public void operationsDisabled(DsoClusterEvent dsoclusterevent) {
//;
}

public void operationsEnabled(DsoClusterEvent dsoclusterevent) {
//
}


And then call registerListener from the main() function:


App app = new App();
app.registerListener();
app.addMessage("Hello, world");


now run the app with the following command to see the node events


mvn clean install
mvn tc:run


Much much more simpler then registering ShutDown hooks. The project is here

Terracotta Integration Module with Maven

TIMs or Terracotta Integration Module is a way to package classes and configuration for certain interfaces and products so that you can add terracotta clustering without much effort. For example, we have a TIM for ehcache so that anyone who uses ehcache can get terracotta clustering just by including the TIM http://forge.terracotta.org/releases/projects/tim-ehcache/

Sometimes using archetypes in maven is a bit unweilding, so first we are going to check out the correct versions and make sure everything is in place.

Let's check out the maven-archetype-plugin first:


svn co http://svn.apache.org/repos/asf/maven/archetype/tags/maven-archetype-2.0-alpha-4/ maven-archetype-2.0-alpha-4
cd maven-archetype-2.0-alpha-4
mvn clean install


Now we want to checkout and install the tim-archetype plugin, is is what builds the directory structure and configuration for our TIM project.


svn co http://svn.terracotta.org/svn/forge/projects/tim-archetype/trunk/ tim-archetype
cd tim-archetype
mvn clean install


To create a TIM project you need to run the following command:


mvn org.apache.maven.plugins:maven-archetype-plugin:2.0-alpha-4:generate \
-DarchetypeGroupId=org.terracotta.maven.archetypes \
-DarchetypeArtifactId=tim-archetype \
-DarchetypeVersion=1.5.0-SNAPSHOT \
-DremoteRepositories=http://www.terracotta.org/download/reflector/maven2 \
-DgroupId=org.terracotta.modules.memcached \
-DartifactId=tim-memcached \
-Dversion=1.0.0



this will prompt you to run the following command:


[INFO] Using property: groupId = org.terracotta.modules.memcached
[INFO] Using property: artifactId = tim-memcached
[INFO] Using property: version = 1.0.0
Define value for property 'package': org.terracotta.modules.memcached.tim-memcached: :


Hit enter or type in the package name you desire and hit enter.


Define value for property 'package': org.terracotta.modules.memcached.tim-memcached: :
Confirm properties configuration:
groupId: org.terracotta.modules.memcached
artifactId: tim-memcached
version: 1.0.0
package: org.terracotta.modules.memcached.tim-memcached
Y: :


Type in 'Y' to finish creating a newly minted TIM.

All set to go!

Back to Basics

When doing a little refactoring on terracotta server code, I was reminded on why the basics ALWAYS matters.

The server side representation of both ArrayList and LinkedList mapped to a class called ListManagedObjectState. The underlying data structure for this class was a ArrayList. We got around to mapping LinkedList to a LinkedListManagedObjectState where the underlying data structure for that class is a LinkedList.

I wrote a little test and the following where the results:


ArrayList.add for 50000 objects took 424 ms.
LinkedList.add for 50000 objects took 198 ms.
ArrayList.addFirst for 50000 objects took 906 ms.
LinkedList.addFirst for 50000 objects took 109 ms.
ArrayList.addAt for 50000 took objects 105 ms.
LinkedList.addAt for 50000 took objects 132 ms.
ArrayList.clear for 50000 took objects 127 ms.
LinkedList.clear for 50000 took objects 49 ms.
ArrayList.removeFirst for 50000 took objects 847 ms.
LinkedList.removeFirst for 50000 took objects 70 ms.
ArrayList.remove for 50000 took objects 865 ms.
LinkedList.remove for 50000 took objects 68 ms.


If your mutating a lot more then accessing, then LinkedList is the way to go.
Basics do matter!

DSO Garbage Collection: YoungGen

There are certain apps when clustered using Terracotta that generates short-lived DSO objects and lots of them. Which results in a lot of DSO garbage!

Taking the existing DSO garbage collector and running it more often will collect these short-lived objects, but the MARK stage could take a long time. For example, you'll have a Terracotta server array that has DSO objects on the server's cache as well as DSO objects faulted to disk. When the DSO garbage collector runs, it will fault in objects from the disk during the MARK stage. This may take awhile to run.

The DSO objects that are short-lived are probably in the server's cache. This is where YoungGen comes in. It's the same algorithm, except for the set of objects we are considering as GC Candidates. Instead of considering all the objects in the system and removing objects that are unreachable. We only consider the objects on the server's cache and remove objects that we can definitely determine as garbage.

DSO objects that cannot be collected in YoungGen is the following:

1. Roots
2. DSO Objects that in the server's cache that is being references by DSO Objects faulted to disk. This is marked in a subset called RememberMe set
3. Objects that are in the cache, but never been flushed to disk (i.e. really new objects).

enable the following properties and see if YoungGen DSO Garbage Collection works for you:


l2.objectmanager.dgc.young.enabled = true
l2.objectmanager.dgc.young.frequencyInMillis = 180000


Terracotta DSO GC Algorithm in a nutshell

There maybe a few of you out there (when I mean few, hopefully a few million :)) that maybe interested in how Terracotta collects its Distributed Objects that is no longer in use. Ok here goes...

Objects that should be collected by DSO Garbage Collection are objects which are no longer referenced by any client (meaning the object has already been GCed locally on every client that held it). So now that the clients have no more use for this object, its up to the server array to dispose of it (i.e. remove it from the server array cache or terraocotta's persistent storage).

When DGC kicks in..

1. It starts monitoring for new objects being created and keep a set of those. This is used later to rescue GC candidates.

2. It does its initial mark, which is to get all the ObjectIDs in the system. And also our roots in the system. We remove the roots and well as object reachable by roots form GC Candidates set. The objects that are left are objects that are no longer referenced by anyone

3. Remember when monitoring new object references was turned on? Now new referenced objects that appeared in our GC Candidate set can be removed from that set. This means there were new references to those objects while our first marking was going on, so they need to be rescued and not collected anymore. If after this rescue there are no GC candidates, then GC is complete.

4. If there are still GC Candidates left, then the system is paused. For us this means pausing our ObjectManager so that new lookup or create requests
goes pending (i.e no new references are created). We now do another rescue and produce our final GC Candidate set for this GC Cycle. Then new reference monitoring is turned off.

5. finally we have a have a set to delete and pass that list onto a different thread that will delete the GC Candidates from the ObjectManager in batches.

Since the delete is staged on a different thread (SEDA arch). It is possible for another GC Cycle to kick in before all the delete request complete, which can in turn create more objects to delete and life goes on...

Up next: Young GC Candidates....