Low GC memory buffers

In-memory, bounded-size queue data structure that aims at minimizing garbage collection overhead (especially latency)

License

License

GroupId

GroupId

com.fasterxml.util
ArtifactId

ArtifactId

low-gc-membuffers
Last Version

Last Version

1.2.0
Release Date

Release Date

Type

Type

bundle
Description

Description

Low GC memory buffers
In-memory, bounded-size queue data structure that aims at minimizing garbage collection overhead (especially latency)
Project URL

Project URL

http://github.com/cowtowncoder/low-gc-membuffers
Project Organization

Project Organization

fasterxml.com
Source Code Management

Source Code Management

http://github.com/cowtowncoder/low-gc-membuffers

Download low-gc-membuffers

Dependencies

test (1)

Group / Artifact Type Version
junit : junit jar 4.12

Project Modules

There are no modules declared in this project.

Overview

This project aims at creating a simple efficient building block for "Big Data" libraries, applications and frameworks; thing that can be used as an in-memory, bounded queue with opaque values (sequence of JDK primitive values): insertions at tail, removal from head, single entry peeks), and that has minimal garbage collection overhead. Insertions and removals are as individual entries, which are sub-sequences of the full buffer.

GC overhead minimization is achieved by use of direct ByteBuffers (memory allocated outside of GC-prone heap); and bounded nature by only supporting storage of simple primitive value (byte, `long') sequences where size is explicitly known.

Conceptually memory buffers are just simple circular buffers (ring buffers) that hold a sequence of primitive values, bit like arrays, but in a way that allows dynamic automatic resizings of the underlying storage. Library supports efficient reusing and sharing of underlying segments for sets of buffers, although for many use cases a single buffer suffices.

Buffers vary in two dimensions:

  1. Type of primitive value contained: currently byte and long variants are implemente, but others (like int or char) will be easy to add as needed
  2. Whether sequences are "chunky" -- sequences consists of 'chunks' created by distinct appendEntry() calls (and retrieved in exactly same sized chunks with getNextEntry()) -- or "streamy", meaning that values are coalesced and form a logical stream (so multiple appendEntry() calls may be coalesced into just one entry returned by getNextEntry()).

Since Java has no support for "generic primitives", there are separate classes for all combinations. This means that there are currently 4 flavors of buffers:

  • for byte (using MemBuffersForBytes)
  • ChunkyBytesMemBuffer
  • StreamyBytesMemBuffer
  • for long (using MemBuffersForLongs)
  • ChunkyLongsMemBuffer
  • StreamyLongsMemBuffer

Another thing that can vary is the way underlying segments are allocated; default is to use native ("direct") ByteBuffers. But more on this later on.

Licensing

Standard Apache 2.0 license.

Fancier stuff: multiple buffers

Although having individual buffers is useful as is, this is just the beginning. Conceptually library supports "buffer groups", sets of similary-valued buffer instances owned by a single factory (like MemBuffersForBytes) that share same segment allocator (com.fasterxml.util.membuf.SegmentAllocator). This makes it possible to share set of reusable underlying ByteBuffer instances for buffers in the same group.

This ability to share underlying segments between buffers, with strict memory bounds makes it possible to use library as basic buffer manager; for example to buffer input and/or output of a web server (byte-based "streamy" buffers), or as simplistic event queues (usually using "chunky" buffers).

To have multiple buffer groups simply construct multiple factory instances.

Thread-safety

All pieces are designed to be used by multiple threads (often just 2, producer/consumer), so all access is properly synchronized.

In addition, locking is done using buffer instances, so it may occasionally make sense to synchronize on buffer instance since this allows you to create atomic sequences of operations, like so:

MemBuffersForBytes factory = new MemBuffersForBytes(...);
ChunkyBytesMemBuffer buffer = factory.createChunkyBuffer(...);
synchronized (buffer) {
  // read latest, add right back:
  byte[] msg = buffer.getNextEntry();
  buffer.appendEntry(msg);
}

or similarly if you need to read a sequence of entries as atomic unit.

Status

Project has been used by multiple production systems (by multiple companies) since 2012, and by now has proven stable and performant for expected use cases. As such it is considered production ready: the official 1.0 version was released in October 2013.

The first accessible project that uses it is Arecibo, a metrics collection, aggregation and visualization.

Companies that use this library for production systems include:

Usage

Getting it

To use with Maven, add:

<dependency>
  <groupId>com.fasterxml.util</groupId>
  <artifactId>low-gc-membuffers</artifactId>
  <version>1.1.1</version>
</dependency>

For downloadables, javadocs check out Wiki.

Start with a factory

Exact factory to use depends on value type: here we assume you are looking for byte-based buffers. If so, you will use MemBuffersForBytes. This object can be viewed as container and factory of actual buffers (ChunkyBytesMemBuffer or StreamyBytesMemBuffer). To construct one, you need to specify amount of memory to use, as well as how memory should be sliced: so, for example:

MemBuffersForBytes factory = new MemBuffersForBytes(30 * 1024, 2, 11);

would create instance that allocates at least 2 (and at most 11) segments (which wrap direct ByteBuffer instances) with size of 30 kB: that is, has memory usage between 60 and 330 kilobytes. The segments are then used by actual buffer instances (more on this in a bit)

So how do you choose parameters? Smaller the segments, more granular is memory allocation, which can mean more efficient memory use (since overhead is bounded to at most 1 segment-full per active buffer). But it also increases number of segment instances, possibly increasing fragmentation and adding overhead.

Note that you can create multiple instances of MemBuffers[Type] factories, if you want to have more control over how pool of segments is allocated amongst individual buffers.

Detour: allocating underlying storage segments

By default segments are allocated as ByteBuffers (or typed sub-types for longs and so on). But this behavior can be changed by passing alternate SegmentAllocator instances.

For example, if you instead wanted to use in-heap segments stored as basic byte arrays (byte[]), you could do this by:

MemBuffersForBytes factory = new MemBuffersForBytes(
  ArrayBytesSegment.allocator(30 * 1024, 2, 11));

or to use non-direct ByteBuffers:

MemBuffersForBytes factory = new MemBuffersForBytes(
  ByteBufferBytesSegment.allocator(30 * 1024, 2, 11, false));

Note that SegmentAllocator instances are implemented as inner classes of matching segment type, that is as ArrayBytesSegment.Allocator and ByteBufferBytesSegment.Allocator.

Also note that neither Allocators nor MemBuffers keep track of underlying segments. What this means it that buffers MUST be closed (explicitly, or indirectly by using wrappers) to make sure segments are released for reuse.

Create individual buffers, MemBuffer

Actual buffers are then allocated using

ChunkyBytesMemBuffer items = bufs.createChunkyBuffer(2, 5);

which would indicate that this buffer will hold on to at least 2 segments (i.e. about 60kB raw storage) and use at most 5 (so max usage of 150kB). Due to circular buffer style of allocation, at least 'segments - 1' amount of memory will be available for actual queue (i.e. guaranteed space of 120kB; that is, up to one segment may be temporarily unavailable depending on pattern of append/remove operations.

And start buffering/unbuffering

To append entries, you use:

byte[] dataEntry = ...; // serialize from, say, JSON
items.appendEntry(dataEntry);

or, if you don't want an exception if there is no more room:

if (!items.tryAppendEntry(dataEntry)) {
   // recover? Drop entry? Log?
}

and to pop entries:

byte[] next = items.getNextEntry(); // blocks if nothing available
// or:
next = items.getNextEntryIfAvailable();
if (next == null) { // nothing yet available
    //...
}
// or:
next = items.getNextEntry(1000L); // block for at most 1 second before giving up

And make sure that...

You '''always close buffers''' when you are done with them -- otherwise underlying segments may be leaked. This because buffers are only objects that keep track of segments; and nothing keeps track of MemBuffer instances created -- this is intentional, as synchronization otherwise needed is very expensive from concurrency perspective.

Note that version 0.9.1 allows use of MemBufferDecorator instances, which makes it possible to build wrappers that can implement simple auto-closing of buffers.

Statistics, anyone?

Finally, you can also obtain various statistics of buffer instances:

int entries = items.getEntryCount(); // how many available for getting?
int segmentsInUse = items.getSegmentCount(); // nr of internal segments
long maxFree = items.getMaximumAvailableSpace(); // approximate free space
long payload = items.getTotalPayloadLength(); // how much used by data?

Download

Check out Wiki for downloads, Javadocs etc.

Known/potential problems

Default (and currently only) buffer implementation uses direct ByteBuffers, and amount of memory that can be allocated is limited by JVM option -XX:MaxDirectMemorySize, which by default has relatively low size of 64megs. To increase this setting, add setting like:

-XX:MaxDirectMemorySize=512m

otherwise you are likely to hit an OutOfMemoryError when using larger buffers.

Future ideas

Here are some improvement ideas:

  • "Slab" allocation (issue #14): allow initial allocation of a longer off-heap memory segment, with size of N segments: this "slab" will be fixed and NOT dynamically allocated or freed; segments will be sub-allocated as needed.
    • Main benefit is reduced need for actually memory management (no per-operation malloc or free)
    • Adds fixed overhead: slab size needs to be balanced with costs
    • Segments from slabs allocated before dynamic segments, as they do not incur additional allocation or memory usage cost (due to fixed default overhead)
  • Expose streamy byte buffers as InputStream (issue #19).
    • Would need to choose what happens with end-of-input: snapshot (expose current and as EOF) vs blocking (works like pipe)

Versions

Version
1.2.0
1.1.1
1.1.0
1.0.0
0.9.3
0.9.2
0.9.1
0.9.0
0.8.1
0.8.0
0.7.0