This guide will give you information and tweaks about tuning Infinispan performance (both server and library mode).

1. Capacity planning

Data in Infinispan is either stored as plain Java objects or in a serialized form, depending on operating mode (embedded or server) or on specific configuration options store-as-binary. Data size can be estimated using sophisticated tools like Java Object Layout and the total amount of required memory can be roughly estimated using the following formulas:

Total Data Set in library mode
\$Total Data Set = Number Of Entries * (Key Size + Value Size + 200 b (Overhead))\$
Total Data Set in server mode
\$Total Data Set = Number Of Entries * (Serialized Key Size + Serialized Value Size + 200 b (Overhead))\$

Term overhead is used here as an average amount of additional memory (e.g. expiration or eviction data) needed for storing an Entry in a Cache.

In case of Local or Replicated mode, all data needs to fit in memory, so calculating the amount of required memory is trivial.

Calculating memory requirements for Distributed mode is slightly more complicated and requires using the following:

Required memory in Distributed mode
\$Required Memo ry = Total Data Set*(Node Failures + 2)/(Nodes - Node Failures)\$

Where:

  • Total Data Set - Estimated size of all data

  • Nodes - The number of nodes in the cluster

  • Node Failures - Number of possible failures (also number of owners - 1)

Calculated amount of memory should be used for setting Xmx and Xms parameters.

JVM as well as Infinispan require additional memory for other tasks like searches, allocating network buffers etc. It is advised to allocate no more than 50% of memory with living data when using Infinispan solely as a caching data store, and no more than 33% of memory with living data when using Infinispan to store and analyze the data using querying, distributed execution or distributed streams.

When considering large heaps, make sure there’s enough CPU to perform garbage collection efficiently.

2. Java Virtual Machine settings

Java Virtual Machine tuning might be divided into sections like memory or GC. Below is a list of helpful configuration parameters and a guide how to adjust them.

2.1. Memory settings

Adjusting memory size is one of the most crucial step in Infinispan tuning. The most commonly used JVM flags are:

  • -Xms - Defines the minimum heap size allowed.

  • -Xmx - Defines the maximum heap size allowed.

  • -Xmn - Defines the minimum and maximum value for the young generation.

  • -XX:NewRatio - Define the ratio between young and old generations. Should not be used if -Xmn is enabled.

Using Xms equal to Xmx will prevent JVM from dynamically sizing memory and might decrease GC pauses caused by resizing. It is a good practice to specify Xmn parameter. This guaranteed proper behavior during load peak (in such case Infinispan generates lots of small, short living objects).

2.2. Garbage collection

The main goal is to minimize the amount of time when JVM is paused. Having said that, CMS is a suggested GC for Infinispan applications.

The most frequently used JVM flags are:

  • -XX:MaxGCPauseMillis - Sets a target for the maximum GC pause time. Should be tuned to meet the SLA.

  • -XX:+UseConcMarkSweepGC - Enables usage of the CMS collector.

  • -XX:+CMSClassUnloadingEnabled - Allows class unloading when the CMS collector is enabled.

  • -XX:+UseParNewGC - Utilize a parallel collector for the young generation. This parameter minimizes pausing by using multiple collection threads in parallel.

  • -XX:+DisableExplicitGC - Prevent explicit garbage collections.

  • -XX:+UseG1GC - Turn on G1 Garbage Collector.

2.3. Other settings

There are two additional parameters which are suggested to be used:

  • -server - Enables server mode for the JVM.

  • -XX:+ UseLargePages - Instructs the JVM to allocate memory in Large Pages. These pages must be configured at the OS level for this parameter to function successfully.

2.4. Example configuration

In most of the cases we suggest using CMS. However when using the latest JVM, G1 might perform slightly better.

32GB JVM
-server
-Xmx32G
-Xms32G
-Xmn8G
-XX:+UseLargePages
-XX:+UseConcMarkSweepGC
-XX:+UseParNewGC
-XX:+DisableExplicitGC
32GB JVM with G1 Garbage Collector
-server
-Xmx32G
-Xms32G
-Xmn8G
-XX:+UseG1GC

3. Network configuration

Infinispan uses TCP/IP for sending packets over the network (for both cluster communication when using TCP stack or when communication with Hot Rod clients)

In order to achieve the best results, it is recommended to increase TCP send and receive window size (refer to you OS manual for instructions). The recommended values are:

  • send window size - 640 KB

  • receive window size - 25 MB

4. Number of threads (Server mode only)

Hot Rod Server uses worker threads which are activated by a client’s requests. It’s important to match the number of worker threads to the number of concurrent client requests:

Hot Rod Server worker thread pool size
<hotrod-connector socket-binding="hotrod" cache-container="local" worker-threads="200">
   <!-- Additional configuration here -->
</hotrod-connector>

5. Cache Store performance

In order to achieve the best performance, please follow the recommendations below when using Cache Stores:

  • Use async mode (write-behind) if possible

  • Prevent cache misses by preloading data

  • For JDBC Cache Store:

    • Use indexes on id column to prevent table scans

    • Use PRIMARY_KEY on id column

    • Configure batch-size, fetch-size, etc

6. Hints for program developers

There are also several hints for developers which can be easily applied to the client application and will boost up the performance.

6.1. Ignore return values

When you’re not interested in returning value of the #put(k, v) or #remove(k) method, use Flag.IGNORE_RETURN_VALUES flag as shown below:

Using Flag.IGNORE_RETURN_VALUES
Cache noPreviousValueCache = cache.getAdvancedCache().withFlags(Flag.IGNORE_RETURN_VALUES);
noPreviousValueCache.put(k, v);

It is also possible to set this flag using ConfigurationBuilder

Using ConfigurationBuilder settings
ConfigurationBuilder cb = new ConfigurationBuilder();
cb.unsafe().unreliableReturnValues(true);

6.2. Use Externalizer for marshalling

Infinispan uses JBoss Marshalling to transfer objects over the wire. The most efficient way to marshall user data is to provide an AdvancedExternalizer. This solutions prevents JBoss Marshalling from sending class name over the network and allows to save some bandwidth:

User entity with Externalizer
import org.infinispan.marshall.AdvancedExternalizer;

public class Book {

   final String name;
   final String author;

   public Book(String name, String author) {
      this.name = name;
      this.author = author;
   }

   public static class BookExternalizer
            implements AdvancedExternalizer<Book> {

      @Override
      public void writeObject(ObjectOutput output, Book book)
            throws IOException {
         output.writeObject(book.name);
         output.writeObject(book.author);
      }

      @Override
      public Person readObject(ObjectInput input)
            throws IOException, ClassNotFoundException {
         return new Person((String) input.readObject(), (String) input.readObject());
      }

      @Override
      public Set<Class<? extends Book>> getTypeClasses() {
         return Util.<Class<? extends Book>>asSet(Book.class);
      }

      @Override
      public Integer getId() {
         return 2345;
      }
   }
}

The Externalizer must be registered in cache configuration. See configuration examples below:

Adding Externalizer using XML
<cache-container>
   <serialization>
      <advanced-externalizer class="Book$BookExternalizer"/>
   </serialization>
</cache-container>
Adding Externalizer using Java
GlobalConfigurationBuilder builder = ...
builder.serialization().addAdvancedExternalizer(new Book.BookExternalizer());

For more information, please refer to the User’s Guide.

6.3. Storing Strings efficiently

If your strings are mostly ASCII, convert them to UTF-8 and store them as byte[]:

  • Using String#getBytes("UTF-8") allows to decrease size of the object

  • Consider using G1 GC with additional JVM flag -XX:+UseStringDeduplication. This allows to decrease memory footprint (see JEP 192 for details).

6.4. Use simple cache for local caches

When you don’t need the full feature set of caches, you can set local cache to "simple" mode and achieve non-trivial speedup while still using Infinispan API.

This is an example comparison of the difference, randomly reading/writing into cache with 2048 entries as executed on 2x8-core Intel® Xeon® CPU E5-2640 v3 @ 2.60GHz:

Table 1. Number of operations per second (± std. dev.)
Cache type single-threaded cache.get(…​) single-threaded cache.put(…​) 32 threads cache.get(…​) 32 threads cache.put(…​)

Local cache

14,321,510 ± 260,807

1,141,168 ± 6,079

236,644,227 ± 2,657,918

2,287,708 ± 100,236

Simple cache

38,144,468 ± 575,420

11,706,053 ± 92,515

836,510,727 ± 3,176,794

47,971,836 ± 1,125,298

CHM

60,592,770 ± 924,368

23,533,141 ± 98,632

1,369,521,754 ± 4,919,753

75,839,121 ± 3,319,835

The CHM shows comparison for ConcurrentHashMap from JSR-166 with pluggable equality/hashCode function, which is used as the underlying storage in Infinispan.

Even though we use JMH to prevent some common pitfals of microbenchmarking, consider these results only aproximative. Your mileage may vary.

Please consult User’s Guide for the limitations of simple cache and configuration.