The Things Network HTTP Integration Part13

Connection multiplexing

For the Proof of Concept(PoC) I had used a cache to store Azure IoT Hub connections to reduce the number of calls to the Device Provisioning Service(DPS).

Number of connections with no pooling

When stress testing with 1000’s of devices my program hit the host connection limit so I enabled Advanced Message Queuing Protocol(AMQP) connection pooling.

return DeviceClient.Create(result.AssignedHub,
                  authentication,
                  new ITransportSettings[]
                  {
                     new AmqpTransportSettings(TransportType.Amqp_Tcp_Only)
                     {
                        PrefetchCount = 0,
                        AmqpConnectionPoolSettings = new AmqpConnectionPoolSettings()
                        {
                           Pooling = true,
                        }
                     }
                  }
               );

My first attempt failed as I hadn’t configured “TransportType.Amqp_Tcp_Only” which would have allowed the AMQP implementation to fallback to other protocols which don’t support pooling.

Exception caused by not using TransportType.Amqp_Tcp_Only

I then deployed the updated code and ran my 1000 device stress test (note the different x axis scales)

Number of connections with pooling

This confirmed what I found in the Azure.AMQP source code

/// <summary>
/// The default size of the pool
/// </summary>
/// <remarks>
/// Allows up to 100,000 devices
/// </remarks>
/// private const uint DefaultPoolSize = 100;

The Things Network HTTP Integration Part12

Removing the DIY cache

For the Proof of Concept(PoC) I had written a simple cache using a ConcurrentDictionary to store Azure IoT Hub connections to reduce the number of calls to the Device Provisioning Service(DPS).

Device Provisioning Service calls in stress test

For a PoC the DIY cache was ok but I wanted to replace it with something more robust like the .Net ObjectCache which is in the System.Runtime.Caching namespace.

I started by replacing the ConcurrentDictionary declaration

static readonly ConcurrentDictionary<string, DeviceClient> DeviceClients = new ConcurrentDictionary<string, DeviceClient>();
     

With an ObjectCache declaration.

static readonly ObjectCache DeviceClients = MemoryCache.Default;
  

Then, where there were compiler errors I updated the method call.

// See if the device has already been provisioned or is being provisioned on another thread.
if (DeviceClients.Add(registrationId, deviceContext, cacheItemPolicy))
{
   log.LogInformation("RegID:{registrationId} Device provisioning start", registrationId);
...

One difference I found was that ObjectCache throws an exception if the value is null. I was using a null value to indicate that the Device Provisioning Service(DPS) process had been initiated on another thread and was underway.

I have been planning to add support for downlink messages so I added a new class to store the uplink (Azure IoT Hub DeviceClient) and downlink ( downlink_url in the uplink message) details.

 public class DeviceContext
   {
      public DeviceClient Uplink { get; set; }
      public Uri Downlink { get; set; }
   }

For the first version the only functionality I’m using is sliding expiration which is set to one day

CacheItemPolicy cacheItemPolicy = new CacheItemPolicy()
{
   SlidingExpiration = new TimeSpan(1, 0, 0, 0),
   //RemovedCallback
};

DeviceContext deviceContext = new DeviceContext()
{
   Uplink = null,
   Downlink = new Uri(payload.DownlinkUrl)
};

I didn’t have to make many changes and I’ll double check my implementation in the next round of stress and soak testing.