I had a customer streaming messages at a high-rate (up to 2000 msg/s - 1KB each) from a protocol translator running on an x86 industrial PC to a cloud-based Mosquito MQTT broker.
That edge device evolved quickly into a more capable and secure intelligent edge solution thanks to Azure IoT Edge and Azure IoT Hub, adding device provisioning (secured with an HW-based identity) and device management capabilities on top of a bi-directional communication, along with the deployment, execution, and monitoring of other edge workloads in addition to a containerized version of the original MQTT protocol translator.
The performance requirement did not change though: the protocol translator (running now as an Azure IoT Edge module) still had to ingest and deliver to the IoT Hub up to 2000 msg/s (1KB each), with a minimum latency.
Is it feasible? Can an IoT Edge solution stream 2000 msg/s or even higher rates? What’s the upper limit? How to minimize the latency?
This blog post will guide you through a detailed analysis of the pitfalls and bottlenecks when dealing with high-rate streams, to eventually show you how to optimize your IoT Edge solution and meet and exceed your performance requirements in terms of rate, throughput, and latency.
The Azure IoT Edge runtime includes a module named edgeHub, acting as a local proxy for the IoT Hub and as a message broker for the connected devices and modules.
The edgeHub supports extended offline operation: if the connection to the IoT Hub is lost, edgeHub saves messages and twin updates in a local message queue (aka “store&forward” queue). Once the connection is re-established, it synchronizes all the data with the IoT Hub.
The environment variable “UsePersistentStorage” controls whether the message queue is:
When persisted on disk (default), the location of the queue will be:
The size of the queue is not capped, and it will grow as long as the device has storage capacity.
When dealing with a high message rate over a long period, the queue size could easily exceed the available storage capacity and cause the OS crash.
Binding the edgeHub folder to a dedicated partition, or even a dedicated disk if available, would protect the OS from uncontrolled growth of the edgeHub queue.
If the “DATA” partition (or disk) runs out of space:
Do size the partition for the worst-case or reduce the Time-To-Live (TTL).
I will let you judge what’s the worst case in your scenario. But the very worst case is total disconnection for the entire TTL. During the TTL (which is 7200 s = 2 hrs. by default), the queue will accumulate all the incoming messages at a given rate and size. And be aware that the edgeHub keeps one queue per endpoint and per priority.
An estimation of the queue size on disk would be:
And if you do the math, a “high” rate of 2000 [msg/s] with 1 [KB/msg] could easily consume almost 15 GBs of storage within 2 hrs.
But even a “low” 100 [msg/s] rate you could easily consume up to 1GB, which would be an issue on constrained devices with an embedded storage of few GBs.
Then, to keep the disk consumption under control:
If you keep the queue disk consumption under control with proper estimation and sizing, you don’t need to bind it to a dedicated partition/disk. But it’s an extra-precaution that comes with almost no effort.
Btw: If you considered setting UsePersistentStorage=false to store the queue in memory, you may realize now that the amount of RAM needed would make it an expensive option if compared to disk or non-viable at all. Moreover, such an in-memory store would NOT be resilient to unexpected crashes or reboots (as the “EnableNonPersistentStorageBackup” can backup and restore the in-memory queue only when you go through a graceful shutdown and reboot).
What happens to expired messages?
Expired messages are removed every 30 minutes by default, but you can tune that interval using this environment variable MessageCleanupIntervalSecs:
If you use different priorities (LINK), do set “CheckEntireQueueOnCleanup”=true to force a deep clean-up and make sure that all expired messages are removed, regardless of the priority.
Why? The edgeHub keeps one queue per endpoint and per priority (but not per TTL)
If you have 2 routes with the same endpoint and priority but different TTL, those messages will be put in the same queue. In that case, it is possible that the messages with different TTLs are interleaved in that queue. When the cleanup processor runs, by default, it checks the messages on the head of each queue to see if they are expired. It does not check the entire queue, to keep the cleanup process as light as possible. If you want it to clean up all messages with an expired TTL, you can set the flag CheckEntireQueueOnCleanup to true.
Now that you have the disk consumption of your edgeHub queue under control, it’s a good practice to keep it monitored using the edgeAgent and edgeHub built-in metrics and the Azure Monitor integration.
The “edgeAgent_available_disk_space_bytes” reports the amount of space left on the disk.
…but there’s another metric you should pay attention to, which is counting the number of non-expired messages still in the queue (i.e. not delivered yet to the destination endpoint):
That “edgehub_queue_length” is a revelation, and it explains how the latency relates to the rates. But to understand it, we must measure the message rate along the pipeline first.
I developed IotEdgePerf, a simple framework including:
Further instructions in the IotEdgePerf GitHub repo.
I deployed the IotEdgePerf transmitter module to an IoT Edge 1.2 (instructions here) running on a DS2-v2 VM and connected to a 1xS3 IoT Hub. I launched the test from the console app as follows:
dotnet run -- \ --payload-length=1024 \ --burst-length=50000 \ --target-rate=2000
Here are the results:
As anticipated, the “edgehub_queue_length” explains why we have the latency. Let’s have a look at it using Log Analytics:
Let’s correlate the queue length with the transmission burst: as the queue is a FIFO (First-In-First-Out), the last message produced by the transmitter is the last message ingested by the IoT Hub. Looking at “edgehub_queue_length” data, the latency on the last message is 42 seconds.
How does the queue’s growth and degrowth slopes and maximum value relate to the message rate?
which is in line with what you would expect from a queue, where the growth rate:
The consistency among the different measurements (on the message rate, queue growth/degrowth and latency) proves that the methodology and tools are correct.
Using some simple math, we can express the latency as:
where N is the number of messages.
If we apply that equation to the numbers we measured, again we can get a perfect match:
The latency will be minimum when rateOUT = rateIN, i.e. when the upstream rate equals the source output rate, and the queue does not accumulate messages. This is quite an obvious outcome, but now you have a methodology and tools to measure and relate to each other the rates, the latency, and the queue length (and the disk consumption as well).
Let’s go back to the original goal of a sustained 2000 msg/s rate delivered upstream with minimum latency. We are now able to measure both the source output and the upstream rate, and to tell what’s the performance gap we must fill to assure minimum latency:
But… how to fill that gap? What’s the bottleneck? Do I need a faster CPU or more cores? More RAM? Or the networking is the bottleneck? Or are we hitting some throttling limits on the IoT Hub?
Let’s try with a more “powerful” hardware.
Even if IoT Edge will usually run on a physical HW, let’s use IoT Edge on an Azure VMs, which provide a convenient way of testing different sizes in a repeatable way and compare the results consistently.
1 vcpu / 3.5 GB / ~20
900 ÷ 1300
500 ÷ 600
2 vcpu / 7 GB / ~40
4 vcpu / 7 GB / ~75
8 vcpu / 14 GB / ~140
16 vcpu / 56 GB / ~300
(test conditions: 1xS3 IoT Hub unit, this C# transmitter module, 300K msg, msg size 1KB)
Scaling UP from DS1 to DS5, the source rate increases by around ~50% from a DS1 to DS5… which is peanuts if we consider that a DS5 performs ~15x better (scores here) and costs ~16x (prices here) than a DS1.
Even more interestingly, the upstream rate does not increase, suggesting there’s a weak correlation (or no correlation at all) with the HW specs.
Let’s distribute the source stream across multiple modules in a kind of scaling OUT, and look at the aggregated rate produced by all the modules.
The maximum aggregated source is ~1900 msg/s rate (obtained with N=3 modules), which is higher than the rate of a single module (~1260 msg/s), with a gain of ~50%. However, such improvement is not worth it if we consider the higher complexity of distributing the source stream across multiple source modules.
Interestingly, the upstream rate increases from ~600 to ~1436 msg/s. Why?
The edgeHub can either use the AMQP or the MQTT protocol to communicate upstream with the cloud, independently from protocols used by downstream devices. The AMQP protocol provides multiplexing capabilities that allow the edgeHub for combining multiple downstream logical connections (i.e. the many modules) into a single upstream connection, which is more efficient and leads to the rate improvement that we measured. AMQP is indeed the default upstream protocol and the one I used in this test. This also confirms that the upstream rate is mostly determined by the protocol stack overhead.
Scaling UP from a 1 vCPU machine (DS1) to a 16 vCPU (DS5) machine didn’t help when using a single source module. But what if we have multiple source modules? Would many cores bring an advantage?
Yes. Multiple modules mean multiple docker containers and eventually multiple processes, that will run more efficiently on a multi-core machine.
But is the 3.5x boost worth the 16x price increase of a DS5 vs DS1? No, if you consider that the upstream, again, didn’t increase.
Let’s go back to a single module and increase the message size from 1 KB to 32 KB on a DS1v2 (1 vcpu, 3.5GB). Here are the results:
(test conditions: 1xS3 IoT Hub unit, this C# transmitter module)
The rate decreases from 839msg/s@1KB down to 227msg/s@32KB, but the THROUGHPUT increases from ~0.8 MB/s up to ~7.2MB/s. Such behavior suggests that sending big messages and at a lower rate is more efficient!
How to leverage it?
Let’s assume that the big 16KB message is a batch of 16 x 1 KB messages: it means that the ~4.5MB/s throughput would be equivalent to a message rate of ~4500 msg/s of 1KB each.
Then message rate and throughput is not the same thing. If the transport protocol (and networking) is the bottleneck, higher throughputs can be achieved by sending bigger messages at a lower rate.
How do we implement message batching then?
We have two options:
The default is MaxUpstreamBatchSize = 10, meaning that some batching is already happening under the cover, even if you didn’t realize it. The optimal value for MaxUpstreamBatchSize would be 256 KB / size(msg), as you would want to fit as many small messages as you can in the batch message.
How does it work?
Set the edgeHub RuntimeLogLevel to DEBUG and look for lines containing “obtained next batch for endpoint iothub”:
Looking at the timestamps, you’ll see that messages are collected and sent upstream in a batch every 20ms, suggesting that:
Are the application-level and built-in batching equivalent?
On the UPSTREAM side, you have batch messages (i.e. big and low-rate) in both cases. It means that:
The latter point is quite interesting: as the message batch counts as 1 device-to-cloud operation, the batching helps also reduce the pressure on the d2c throttling limit of the IoT Hub, which is an attention point when dealing with high message rates.
On the SOURCE side, the two approaches are very different:
Eventually, we can state that:
Let’s test that assumption.
On a DS2v2 VM, the max output rate of the source module is ~1100 [msg/s] (with 1KB messages).
As expected, such source rate does not benefit from higher MaxUpstreamBatchSize values, while the upstream does, and it eventually equals the 1100 [msg/s] source rate (hence no latency).
The application-level batching is effective on both the source and upstream throughput.
On a DS1 (1 vCPU, 3.5GB of RAM, which is the smallest size of the DSv2 family) you can achieve:
Is latency always a problem? Not actually. As we saw, the latency is proportional to the number of messages sent (N):
When sending a short burst of messages, the latency may be negligible. On the other hand, on short burst you may want to minimize the transmission duration.
As an example, let’s assume you want to send N=1000 messages, of 1 KB each. Depending on the batching, the transmission duration at the source side will be:
Shortening the transmission duration from a 1.1s down to 0.1s could be critical in battery powered applications, or when the device must send some information upon an unexpected loss of power.
How do other SDKs (Python, Node.js, Java, C) perform in terms of maximum message rate and throughput? The performance gap, if any, would be due to:
As an example of the latter, the JAVA SDK differs from other SDKs as the device/module client embeds a queue and a periodic thread that checks for queued messages. Both the thread period (“receivePeriodInMilliseconds”) and the number of messages de-queued per execution (“SetMaxMessagesSentPerThread”) can be tweaked to increase the output rate.
On the other hand, what is the common denominator to all the SDKs? It’s the transport protocol, which ultimately sets the upper bound of the achievable performance. In this blog post we focused on MQTT, and it would be worth to explore the performance upper bound by using a fast and lightweight MQTT client instead of the SDK. That’s doable and it’s explained here.
A performance comparison among the different Device SDKs as well as a MQTT client will be the topic for a next blog post.
A bash script to spin up a VM with Ubuntu Server 18.04 and a fully provisioned Azure IoT Edge ready-to-go: https://github.com/arlotito/vm-iotedge-provision
A framework and a CLI tool to measure the rate, the throughput and end-to-end latency of an Azure IoT Edge device:
This blog post provided a detailed analysis of the pitfalls and bottlenecks when dealing with high-rate streams, and showed you how to optimize your IoT Edge solution to meet and exceed your performance requirements in terms of rate, throughput, and latency.
On an Azure DS1v2 Virtual Machine, we were able to meet and exceed the original performance target of 2000 msg/s (1KB each) and minimum latency, and we achieved a sustained end-2-end throughput of 3600 KB/s with no latency, or up to 7300 msg/s (1KB each) with some latency.
With the methodology and tools discussed in this blog post, you can assess the performance baseline on your specific platform and eventually optimize it using the built-in or application-level message batching.
In a nutshell:
Special thanks to the Azure IoT Edge team (Venkat Yalla and Varun Puranik), the Industry Solutions team (Simone Banchieri, Stefano Causarano and Franco Salmoiraghi), the IoT CSU (Vitaliy Slepakov and Michiel van Schaik) and Olivier Bloch for the support and the many inspiring conversations.
You must be a registered user to add a comment. If you've already registered, sign in. Otherwise, register and sign in.