Azure Service Bus Performance Tuning

Prerequisites

An environment variable named AzureServiceBus.ConnectionString with the connection string for the Azure Service Bus namespace.

Environment variables named AzureServiceBus.ConnectionString1 and AzureServiceBus.ConnectionString2, each having a different connection string to an Azure Service Bus namespace.

Azure Service Bus Transport

This sample uses the Azure Service Bus Transport (Legacy).

Code walk-through

There are three endpoints in this sample:

• Sender
• Receiver
• SenderReceiver

Sender

Sender sends SomeMessage to Receiver.

public class SomeMessage :
    IMessage
{
}

The sender sends a large number of messages to the receiver and measures how long it takes. The total time is divided by the number of messages sent in order to compute the average throughput.

There are two variations of the Sender which present different ways of sending messages and illustrate what impact they have on performance.

Receiver

Receiver receives SomeMessage and computes how many it was able to receive every second. Again there are various permutations of settings that impact the receive throughput.

SenderReceiver

SenderReceiver is a variation of the Receiver that does not only receive messages, but it also sends messages to a destination queue.

Running the sample

Run without debugging (CTRL + F5) for optimal results.

• Run the Sender standalone to test the send performance, but also to fill up the receive queue.
• Run the Receiver standalone to test the receive performance. If the queue empties, run the sender again.
• Run the SenderReceiver standalone.

Variations

Slow (sequential) sender

In the slow sender scenario:

• Only one factory is defined, this means that only one TCP connection is maintained with the broker.
• Client-side batching of the ASB SDK is turned off, so the SDK will send only one message at a time.

var factories = transport.MessagingFactories();
factories.BatchFlushInterval(TimeSpan.Zero);
factories.NumberOfMessagingFactoriesPerNamespace(1);
transport.NumberOfClientsPerEntity(1);

In this scenario, each send is awaited individually, which prevents batching (even if batching is enabled on the client).

for (var i = 0; i < NumberOfMessages; i++)
{
    Console.WriteLine("Sending a message...");

    // by awaiting each individual send, no client side batching can take place
    // latency is incurred for each send and thus lowest performance possible
    await endpointInstance.Send(new SomeMessage())
        .ConfigureAwait(false);
}

Fast (concurrent) sender

In the fast sender scenario:

• Multiple connections are established by configuring multiple messaging factories.
• Each factory is matched with a sender object (there should be one sender per factory).
• Client-side batching is on, allowing the ASB SDK to send many messages at once.

var factories = transport.MessagingFactories();
factories.BatchFlushInterval(TimeSpan.FromMilliseconds(100));
var totalConcurrency = Environment.ProcessorCount;
factories.NumberOfMessagingFactoriesPerNamespace(totalConcurrency);
transport.NumberOfClientsPerEntity(totalConcurrency);

Individual sends are not awaited, instead the entire batch is awaited. Code execution will continue when all batches are sent.

var tasks = new List<Task>();
for (var i = 0; i < NumberOfMessages; i++)
{
    var task = endpointInstance.Send(new SomeMessage());
    tasks.Add(task);
}

Console.WriteLine("Waiting for completion...");
// by awaiting the sends as one unit, this code allows the ASB SDK's client side batching to kick in and bundle sends
// this results in less latency overhead per individual sends and thus higher performance
await Task.WhenAll(tasks)
    .ConfigureAwait(false);

Slow (sequential) atomic receiver

In the slow atomic receiver scenario:

• Only one factory is defined, this means that only one TCP connection is maintained with the broker.
• Prefetching is turned off, meaning that the receiver will fetch only one message at a time.
• No concurrent operations are allowed; messages are processed one by one.

var queues = transport.Queues();
queues.EnablePartitioning(true);

transport.Transactions(TransportTransactionMode.SendsAtomicWithReceive);
endpointConfiguration.LimitMessageProcessingConcurrencyTo(1);
var receivers = transport.MessageReceivers();
receivers.PrefetchCount(0);

var factories = transport.MessagingFactories();
factories.NumberOfMessagingFactoriesPerNamespace(1);
transport.NumberOfClientsPerEntity(1);

Fast (concurrent) atomic receiver

In the fast atomic receiver scenario:

• Multiple connections are established with the broker.
• Each connection is matched with exactly one receive client.
• Ensures that at least 16 connections are established when using partitioned queues, as each partitioned queue consists of 16 partitions. Depending on the available bandwidth, more can be added until the network saturates.
• Transaction mode AtomicSendsWithReceive will execute complete operations in a serializable transaction, which means one by one. Therefore it is not beneficial to allow high concurrency per receiver. A value of 2 allows one receive to start while another is completed. The more time execution takes, the higher this value can be, but in that scenario one should never use very high values.
• PrefetchCount is set to a relatively low number, since receiver concurrency and throughput are also low.

transport.Transactions(TransportTransactionMode.SendsAtomicWithReceive);

var queues = transport.Queues();
queues.EnablePartitioning(true);

// values 2 and 4 work best, as tx is serializable it makes no sense to allow many concurrent tasks
var perReceiverConcurrency = 2;

// increase number of receivers as much as bandwidth allows
var numberOfReceivers = 16;

var globalConcurrency = numberOfReceivers * perReceiverConcurrency;

endpointConfiguration.LimitMessageProcessingConcurrencyTo(globalConcurrency);
var receivers = transport.MessageReceivers();
receivers.PrefetchCount(20);

var factories = transport.MessagingFactories();
factories.NumberOfMessagingFactoriesPerNamespace(numberOfReceivers);
transport.NumberOfClientsPerEntity(numberOfReceivers);

Fast (concurrent) atomic sender receiver

The fast atomic sender receiver scenario is similar to the fast atomic receiver scenario, except that it establishes twice the number of connections so that the receive clients and send clients can leverage different connections.

transportConfiguration.Transactions(TransportTransactionMode.SendsAtomicWithReceive);

var queues = transportConfiguration.Queues();
queues.EnablePartitioning(true);

// values 2 and 4 work best, as tx is serializable it makes no sense to allow many concurrent tasks
var perReceiverConcurrency = 8;

// increase number of receivers as much as bandwidth allows (probably less than receiver due to send volume)
var numberOfReceivers = 16;

var globalConcurrency = numberOfReceivers * perReceiverConcurrency;

endpointConfiguration.LimitMessageProcessingConcurrencyTo(globalConcurrency);
var receivers = transportConfiguration.MessageReceivers();
receivers.PrefetchCount(20);

var factories = transportConfiguration.MessagingFactories();
factories.NumberOfMessagingFactoriesPerNamespace(numberOfReceivers * 2);
transportConfiguration.NumberOfClientsPerEntity(numberOfReceivers);

Slow (sequential) non-atomic receiver

In the sequential non-atomic receiver scenario:

• Only one factory is defined, this means that only one TCP connection is maintained with the broker.
• Prefetching is turned off, meaning that the receiver will fetch only one message at a time.
• No concurrent operations are allowed; messages are processed one by one.

transport.Transactions(TransportTransactionMode.ReceiveOnly);
var queues = transport.Queues();
queues.EnablePartitioning(true);
endpointConfiguration.LimitMessageProcessingConcurrencyTo(1);
var receivers = transport.MessageReceivers();
receivers.PrefetchCount(0);

var factories = transport.MessagingFactories();
factories.NumberOfMessagingFactoriesPerNamespace(1);
transport.NumberOfClientsPerEntity(1);

Fast (concurrent) non-atomic receiver

In the fast non-atomic receiver scenario:

• Multiple connections are established with the broker.
• Each connection is matched with exactly one receive client.
• Ensures that at least 16 connections are established when using partitioned queues, as each partitioned queue consists of 16 partitions. Depending on available bandwidth configure, more can be added until the network saturates.
• TransactionMode ReceiveOnly can batch message completions, allowing receive operations to work independently. Therefore it is beneficial to allow high concurrency per receiver. For example, 128 operations per receiver if the operation can handle this kind of concurrency. The effect that concurrency level will have on the application depends on the code inside the handler, e.g. a database connection, may not support 128 concurrent operations, while an in-memory operation on a concurrent data structure will benefit from it.
• PrefetchCount is set to a high number as well, usually 1x or 2x the allowed per-receiver concurrency.

transport.Transactions(TransportTransactionMode.ReceiveOnly);

var queues = transport.Queues();
queues.EnablePartitioning(true);

var numberOfCores = Environment.ProcessorCount;

// concurrency allowed
var perReceiverConcurrency = 128;

// increase number of receivers as much as bandwidth allows
var numberOfReceivers = 32;

var globalConcurrency = numberOfReceivers * perReceiverConcurrency;

endpointConfiguration.LimitMessageProcessingConcurrencyTo(globalConcurrency);

// as is prefetching
var receivers = transport.MessageReceivers();
receivers.PrefetchCount(perReceiverConcurrency);

var factories = transport.MessagingFactories();
factories.NumberOfMessagingFactoriesPerNamespace(numberOfReceivers);
transport.NumberOfClientsPerEntity(numberOfReceivers);

Fast (concurrent) non-atomic sender receiver

The fast non-atomic sender receiver scenario is similar to the fast non-atomic receiver scenario except that it establishes twice the number of connections, so that the receive clients and send clients can leverage different connections.

transport.Transactions(TransportTransactionMode.ReceiveOnly);

var queues = transport.Queues();
queues.EnablePartitioning(true);

//lower concurrency if sending more message per receive
var perReceiverConcurrency = 128;

// increase number of receivers as much as bandwidth allows (probably less than receiver due to send volume)
var numberOfReceivers = 16;

var globalConcurrency = numberOfReceivers * perReceiverConcurrency;

endpointConfiguration.LimitMessageProcessingConcurrencyTo(globalConcurrency);
var receivers = transport.MessageReceivers();
receivers.PrefetchCount(perReceiverConcurrency);

var factories = transport.MessagingFactories();
factories.NumberOfMessagingFactoriesPerNamespace(numberOfReceivers * 2);
transport.NumberOfClientsPerEntity(numberOfReceivers);