Channels ▼


ZeroMQ: The Design of Messaging Middleware

Allocating Memory

Assuming that all the infrastructure was initialized and a connection between two endpoints has been established, there's only one thing to allocate when sending a message: the message itself. Thus, to optimize the critical path we had to look into how messages are allocated and passed up and down the stack.

It's common knowledge in the high-performance networking field that the best performance is achieved by carefully balancing the cost of message allocation and the cost of message copying (for example, different handling of small, medium, and large messages). For small messages, copying is much cheaper than allocating memory. It makes sense to allocate no new memory chunks at all and instead to copy the message to preallocated memory whenever needed. For large messages, on the other hand, copying is much more expensive than memory allocation. It makes sense to allocate the message once and pass a pointer to the allocated block, instead of copying the data. This approach is called "zero-copy".

ØMQ handles both cases in a transparent manner. A ØMQ message is represented by an opaque handle. The content of very small messages is encoded directly in the handle. So making a copy of the handle actually copies the message data. When the message is larger, it's allocated in a separate buffer and the handle contains just a pointer to the buffer. Making a copy of the handle doesn't result in copying the message data, which makes sense when the message is megabytes long (Figure 3). It should be noted that in the latter case the buffer is reference-counted so that it can be referenced by multiple handles without the need to copy the data.

Figure 3: Message copying (or not).

Lesson learned: When thinking about performance, don't assume there's a single best solution. It may happen that there are several subclasses of the problem (for example, small messages vs. large messages), each having its own optimal algorithm.


It has already been mentioned that the sheer number of system calls in a messaging system can result in a performance bottleneck. Actually, the problem is much more generic than that. There's a non-trivial performance penalty associated with traversing the call stack and thus, when creating high-performance applications, it's wise to avoid as much stack traversing as possible.

Consider Figure 4. To send four messages, you have to traverse the entire network stack four times (ØMQ, glibc, user/kernel space boundary, TCP implementation, IP implementation, Ethernet layer, the NIC itself and back up the stack again).

Figure 4: Sending four messages.

However, if you decide to join those messages into a single batch, there would be only one traversal of the stack (Figure 5). The impact on message throughput can be overwhelming: up to two orders of magnitude, especially if the messages are small and hundreds of them can be packed into a single batch.

Figure 5: Batching messages.

On the other hand, batching can have negative impact on latency. Let's take, for example, the well-known Nagle's algorithm, as implemented in TCP. It delays the outbound messages for a certain amount of time and merges all the accumulated data into a single packet. Obviously, the end-to-end latency of the first message in the packet is much worse than the latency of the last one. Thus, it's common for applications that need consistently low latency to switch Nagle's algorithm off. It's even common to switch off batching on all levels of the stack (for example, a NIC's interrupt coalescing feature). But no batching means extensive traversing of the stack and results in low message throughput. We seem to be caught in a throughput versus latency dilemma.

ØMQ tries to deliver consistently low latencies combined with high throughput using the following strategy: When message flow is sparse and doesn't exceed the network stack's bandwidth, ØMQ turns all the batching off to improve latency. The trade-off here is somewhat higher CPU usage — we still have to traverse the stack frequently. However, that isn't considered to be a problem in most cases.

When the message rate exceeds the bandwidth of the network stack, the messages have to be queued — stored in memory till the stack is ready to accept them. Queuing means the latency is going to grow. If the message spends one second in the queue, end-to-end latency will be at least one second. What's even worse, as the size of the queue grows, latencies will increase gradually. If the size of the queue is not bound, the latency can exceed any limit.

It has been observed that even though the network stack is tuned for lowest possible latency (Nagle's algorithm switched off, NIC interrupt coalescing turned off, and so on) latencies can still be dismal because of the queuing effect, as described above.

In such situations it makes sense to start batching aggressively. There's nothing to lose as the latencies are already high anyway. On the other hand, aggressive batching improves throughput and can empty the queue of pending messages — which in turn means the latency will gradually drop as the queuing delay decreases. Once there are no outstanding messages in the queue, the batching can be turned off to improve the latency even further.

One additional observation is that the batching should only be done at the topmost level. If the messages are batched there, the lower layers have nothing to batch anyway, and so all the batching algorithms underneath do nothing except introduce additional latency.

Lesson learned: To get optimal throughput combined with optimal response time in an asynchronous system, turn off all the batching algorithms on the low layers of the stack and batch on the topmost level. Batch only when new data are arriving faster than they can be processed.

Architecture Overview

Up to this point I have focused on generic principles that make ØMQ fast. Now, let's have a look at the actual architecture of the system (Figure 6).

Figure 6: ØMQ architecture.

The user interacts with ØMQ using so-called "sockets". They are pretty similar to TCP sockets, the main difference being that each socket can handle communication with multiple peers, a bit like unbound UDP sockets do.

The socket object lives in the user's thread (see the discussion of threading models in the next section). Aside from that, ØMQ is running multiple worker threads that handle the asynchronous part of the communication: reading data from the network, queuing messages, accepting incoming connections, etc.

There are various objects living in the worker threads. Each of these objects is owned by exactly one parent object (ownership is denoted by a simple full line in the diagram). The parent can live in a different thread than the child. Most objects are owned directly by sockets; however, there are couple of cases where an object is owned by an object which is owned by the socket. What we get is a tree of objects, with one such tree per socket. The tree is used during shut down; no object can shut itself down until it closes all its children. This way we can ensure that the shutdown process works as expected; for example, that pending outbound messages are pushed to the network prior to terminating the sending process.

Roughly speaking, there are two kinds of asynchronous objects: objects that are not involved in message passing and those that are. The former have to do mainly with connection management. For example, a TCP listener object listens for incoming TCP connections and creates an engine/session object for each new connection. Similarly, a TCP connector object tries to connect to the TCP peer and when it succeeds it creates an engine/session object to manage the connection. When such connection fails, the connector object tries to re-establish it.

The latter are objects that are handling data transfer itself. These objects are composed of two parts: the session object is responsible for interacting with the ØMQ socket, and the engine object is responsible for communication with the network. There's only one kind of the session object, but there's a different engine type for each underlying protocol ØMQ supports. Thus, we have TCP engines, IPC (inter-process communication) engines, PGM engines (a reliable multicast protocol, see RFC 3208), etc. The set of engines is extensible — in the future we may choose to implement, say, a WebSocket engine or an SCTP engine.

The sessions are exchanging messages with the sockets. There are two directions to pass messages in and each direction is handled by a pipe object. Each pipe is basically a lock-free queue optimized for fast passing of messages between threads.

Finally, there's a context object (discussed in the previous sections but not shown on the diagram) that holds the global state and is accessible by all the sockets and all the asynchronous objects.

Concurrency Model

One of the requirements for ØMQ is to take advantage of multi-core boxes; in other words, to scale the throughput linearly with the number of available CPU cores.

Our previous experience with messaging systems showed that using multiple threads in a classic way (critical sections, semaphores, etc.) doesn't yield much performance improvement. In fact, a multi-threaded version of a messaging system can be slower than a single-threaded version, even if measured on a multi-core box. Individual threads are simply spending too much time waiting for each other while, at the same time, eliciting a lot of context switching that slows the system down.

Given these problems, we decided to go for a different model. The goal was to avoid locking entirely and let each thread run at full speed. The communication between threads was to be provided via asynchronous messages (events) passed between the threads. This, as insiders know, is the classic actor model.

The idea was to launch one worker thread per CPU core — having two threads sharing the same core would only mean a lot of context switching for no particular advantage. Each internal ØMQ object, such as say, a TCP engine, would be tightly bound to a particular worker thread. That, in turn, means that there's no need for critical sections, mutexes, semaphores and the like. Additionally, these ØMQ objects won't be migrated between CPU cores so would thus avoid the negative performance impact of cache pollution (Figure 7).

Figure 7: Multiple worker threads.

This design makes a lot of traditional multi-threading problems disappear. Nevertheless, there's a need to share the worker thread among many objects, which in turn means there needs to be some kind of cooperative multitasking. This means we needed a scheduler; objects need to be event-driven rather than being in control of the entire event loop. That is, we have to take care of arbitrary sequences of events, even very rare ones and we have to make sure that no object holds the CPU for too long; etc.

In short, the whole system had to become fully asynchronous. No object can afford to do a blocking operation, because it would not only block itself but also all the other objects sharing the same worker thread. All objects must become, whether explicitly or implicitly, state machines. With hundreds or thousands of state machines running in parallel you have to take care of all the possible interactions between them and — most importantly — of the shutdown process.

Related Reading

More Insights

Currently we allow the following HTML tags in comments:

Single tags

These tags can be used alone and don't need an ending tag.

<br> Defines a single line break

<hr> Defines a horizontal line

Matching tags

These require an ending tag - e.g. <i>italic text</i>

<a> Defines an anchor

<b> Defines bold text

<big> Defines big text

<blockquote> Defines a long quotation

<caption> Defines a table caption

<cite> Defines a citation

<code> Defines computer code text

<em> Defines emphasized text

<fieldset> Defines a border around elements in a form

<h1> This is heading 1

<h2> This is heading 2

<h3> This is heading 3

<h4> This is heading 4

<h5> This is heading 5

<h6> This is heading 6

<i> Defines italic text

<p> Defines a paragraph

<pre> Defines preformatted text

<q> Defines a short quotation

<samp> Defines sample computer code text

<small> Defines small text

<span> Defines a section in a document

<s> Defines strikethrough text

<strike> Defines strikethrough text

<strong> Defines strong text

<sub> Defines subscripted text

<sup> Defines superscripted text

<u> Defines underlined text

Dr. Dobb's encourages readers to engage in spirited, healthy debate, including taking us to task. However, Dr. Dobb's moderates all comments posted to our site, and reserves the right to modify or remove any content that it determines to be derogatory, offensive, inflammatory, vulgar, irrelevant/off-topic, racist or obvious marketing or spam. Dr. Dobb's further reserves the right to disable the profile of any commenter participating in said activities.

Disqus Tips To upload an avatar photo, first complete your Disqus profile. | View the list of supported HTML tags you can use to style comments. | Please read our commenting policy.