Edge computing describes the movement of computation away from cloud data centers so that it can be closer to instruments, sensors and actuators where it will be run on “small” embedded computers or nearby “micro-datacenters”.  The primary reason to do this is to avoid the network latency in cases where responding to a local event is time critical.  This is clearly the case for robots such as autonomous vehicles, but it is also true of controlling many scientific or industrial apparatuses.  In other cases, privacy concerns can prohibit sending the data over an external network.

We have now entered the age where advances in machine learning has made it possible to infer much more knowledge from a collection of the sensors than was possible a decade ago.    The question we address here is how much deep computational analysis can be moved to the edge and how much of it must remain in the cloud where greater computational resources are available.

The cloud has been where the tech companies have stored and analyze data.  These same tech companies, in partnership with the academic research community, have used that data to drive a revolution in machine learning.  The result has been amazing advances in natural language translation, voice recognition, image analysis and smart digital assistants like Seri, Cortona and Alexa.   Our phones and smart speakers like Amazon Echo operate in close connection with the cloud.   This is clearly the case when the user’s query requires a back-end search engine or database, but it is also true of the speech understanding task.  In the case of Amazon’s Echo, the keyword “Alexa” starts a recording and the recorded message is sent to the Amazon cloud for speech recognition and semantic analysis.   Google cloud, AWS, Azure, Alibaba, Tencent,  Baidu and other public clouds all have on-line machine learning services that can be accessed via APIs from client devises.

While the cloud business is growing and maturing at an increasingly rapid rate, edge computing has emerged as a very hot topic.  There now are two annual research conferences on the subject: the IEEE Service Society International conference on Edge computing and the ACM IEEE Symposium on Edge computing.  Mahadev Satyanarayanan from CMU, in a keynote at the 2017 ACM IEEE Symposium and in the article “The Emergence of Edge Computing” IEEE Computer, Vol. 50, No. 1, January 2017, argues very strongly in favor of a concept called a cloudlet which is a server system very near or collocated with edge devices under its control.   He observes that applications like augmented reality require real-time data analysis and feedback to be usable.  For example, the Microsoft Hololens mixed reality system integrates a powerful 32bit Intel processor with a special graphics and sensor processor.   Charlie Catlett and Peter Beckman from Argonne National Lab have created a very powerful Edge computing platform called Waggle (as part of the Array of Things project) that consists of a custom system management board for keep-alive services and a powerful ODROID multicore processor and a package of instruments that measure Carbon Monoxide, Hydrogen Sulphide, Nitrogen Dioxide, Ozone, Sulfur Dioxide, Air Particles, Physical Shock/Vibration, Magnetic Field, Infrared Light, Ultraviolet Intensity, RMS Sound Level and a video camera.  For privacy reasons the Waggle vision processing must be done completely on the device so that no personal identifying information goes over the network.

Real time computer vision tasks are among the AI challenges that are frequently needed at the edge.  The specific tasks range in complexity from simple object tracking to face and object recognition.   In addition to Hololens and Waggle there are several other small platforms designed to support computer vision at the edge.   As shown in Figure 1, these include the humble RaspberryPi with camera,  the Google vision kit and the AWS DeepLens.

Figure 1.  From the left is a RaspberryPi with an attached camera, ANL Waggle array, the Google AIY vision kit and the AWS DeepLens.

The Pi system is, by far, the least capable with a quad core ARMv7 processor and 1 GB memory.   The Google vision kit has a Raspberry Pi Zero W (single core ARMv7 with 512MB memory) but the real power lies in the Google VisionBonnet which uses a version of the Movidius Myriad 2 vision processing chip which has 12 vector processing units and a dual core risc cpu.   The VisionBonnet runs TensorFlow from a collection of pretrained models.    DeepLens has a 4 megapixel camera, 8 GB memory, 16 GB storage and an intel Atom process and Gen9 graphics engine which supports model built with Amazon SageMaker that is pre-configured to run TensorFlow and Apache MXNet.

As we stated above many applications that run on the edge many must rely on the cloud if only for storing data to be analyzed off-line.   Others, such as many of our phone apps and smart speakers, use the cloud for backend computation and search.   It may be helpful to think of the computational capability of edge devices and the cloud as a single continuum of computational space and an application as an entity that has components distributed over both ends.   In fact, depending upon the circumstances parts of the computation may migrate from the cloud to the device or back to optimize performance.   As illustrated in Figure 2, AWS Greengrass accomplishes some of this by allow you to move Lambda “serverless” functions from the cloud to the device to form a network of long running functions that can interact with instruments and securely invoke AWS services.

Figure 2.   AWS Greengrass allows us to push lambda functions from the cloud to the device and for these functions to communicate seamlessly with the cloud and other functions in other devices.  (Figure from https://aws.amazon.com/greengrass/ )

The Google vision kit is not available yet and DeepLens will ship later in the spring and we will review them when they arrive.   Here we will focus on a few simple experiments with the Raspberry Pi and return to these other devices in a later post.

Deep Learning Models and the Raspberry PI 3.

In a previous post we looked at several computer vision tasks that used the Pi in collaboration with cloud services.   These included simple object tracking and doing optical character recognition and search for information about book covers seen in an image.   In the following paragraphs we will focus on the more complex task of recognizing objects in images and we will try to understand the limitations and advantages of using the cloud as the backend computational resource.

As a benchmark for our experiments we use the Apache MXNet deep learning kit with a model based on the resnet 152-layer neural network that was trained on a  collection  of over a  10 million images and over 11 thousand labels.    We have packaged this MXNet with this model into a Docker container dbgannon/mxnet which we have used for these experiments.  (the details of the python code in the container are in the appendix to this blog.

Note: If you want to run this container and if you have dockerand Jupyter installed you can easily test the model with pictures of your own.  Just download the jupyter notebook send-to-mxnet-container.ipynb and follow the instructions there.

How fast can we do the image analysis (in image frames per second)?

Running the full resnet-152 model on an installed version  of MXNet  on more capable machines (Mac mini and the AWS Deeplearning AMI c5.4xlarge, no GPU)  yields an average performance of about 0.7 frame/sec.     Doing  the same experiment on the same machines, but using the docker container and a local version of the Jupyter notebook driver we see the performance degrade a bit to an average of about 0.69 frame/sec (on a benchmark set of images we described in the next paragraph).    With a GPU one should be able to go about 10 times faster.

For the timing tests we used a set of 20 images from the internet that we grabbed and reduced so they average about 25KB in size.  These are stored in the Edge device.  Loading one of these images takes about the same amount of time as grabbing a frame from the camera and reducing it to the same size.  Two of images from the benchmark set and the analysis output is shown in figure 3 below.

Figure 3. Two of the sample images together with the output analysis and call time.

How can we go faster on the Pi 3?   We are also able to install MXNet on the Pi 3, but it is a non-trivial task as you must build it from the source.   Deployment details are here, however, the resnet 152 model is too large for the 1MB memory of the Pi 3, so we need to find another approach.

The obvious answer is to use a much smaller model such as the Inception 21 layer network which has a model database of only 23MB (vs 310MB  for resnet 152), but it has only 1000 classes vs   the 11,000 of the full rennet 152. We installed Tensorflow on the Pi3. (there are excellent examples of using it for image analysis and recognition provided by Matthew Rubashkin of Silicon Valley Data Science.) We ran the Tensorflow Inception_2015_12_05 which fit in memory on the Pi.  Unfortunately, it was only able to reach 0.48 frames per second on the same image set described above.

To solve the, we need to go to the cloud.    In a manner similar to the Greengrass model,  we will have the Pi3 sample the camera and downsize the image and send it to the cloud for execution.   To test it we ran the MXNet container on a VM in AWS and pointed the Pi camera at various scenes.  The results are shown in Figure 4.

Figure 4.  The result for the toy dinosaur result is as it is logged into the AWS DynamoDB.   With the bottom two images show only the description string.

The output of the model gives us likelihood of various labels.  In a rather simple minded effort to be more conversational we translate the likelihood results as follows.  If a label X is  more than 75% likely  the container returns a value of “This certainly looks like a X”.   If the likely hood is less than less than 35%  it returns “I think this is an X, but I am not sure” (the code is below).    We look at the top 5 likely labels and they are listed in order.

The Pi device pushes jpeg images to AWS S3 as a blob.   It then pushes the metadata about the image (a blob name and time stamp) to the AWS Simple Queue Service.   We modified the MXNet container to wait for something to land in the queue. When this happens, it takes the image meta data and pulls the image from S3 and does the analysis and finally stores the result in an AWS DynamoDB table.

However we can only go as fast as we can push the images  and metadata to the cloud from the Pi device.   With repeated tries we can achieve 6 frames/sec.   To speed up the analysis to match this input stream we spun up a set of analyzers using the AWS Elastic Container Service (ECS).  The final configuration is shown in Figure 5.

Figure 5.   The full Pi 3 to Cloud image recognition architecture. (The test dataset is shown in the tiny pictures in S3)

To conduct the experiments, we included a time stamp from the edge device with the image metadata. When the MXNet container puts the result in the DynamoDB table it includes another timestamp.  This allows us to compute the total time from image capture to result storage for each image in the stream. If the device sends the entire collection as fast as possible then the difference between the earliest recorded time stamp and the most recent gives us a good measure of how long it takes to complete the entire group.

While the Pi device was able to fill S3 and the queue at 6 frames a second having only one MXNet container instance yielded the result that the total throughput was only about 0.4 frames/sec.   The servers used to host the container are relatively small.   However, using the ECS it is trivial to boost the number of servers and instances.  Because of the size of the container instance is so large only one instance can fit on each of the 8 GB servers.   However as shown in Figure 6 we were able to match the device sending throughput with 16 servers/instances.  At this point messages in the queue were being consumed as fast as they were arriving. Using a more powerful device (a laptop with a core I7 processor) to send the images we were about to boost the input end up to just over 7  frames per second and that  was matched with 20 servers/instances.

Figure 6.  Throughput in Frames/second measured from the Pi  device to the final results in the DynamoDB instance.   In the 20 instance case, a faster core I7 laptop was used to send the images.

Final Thoughts

This exercise does not fully explore the utility of AI method deployed at the edge or between the edge and the cloud.  Clearly this type of full object recognition at real-time frame rates is only possible if the edge device has sophisticated accelerator hardware.  On the other hand, there are many simple machine learning models that can be used for more limited applications.   Object motion tracking is one good example.  This can be done in real-time.    This is typically done by comparing a frame to a previous one and looking for the differences.   Suppose you need to invoke fire suppression when a fire is detected.   It would not be had to build a very simple network that can recognize fire but not simple movement of ordinary objects.   Such a network could be invoked whenever movement is detected and if it is fire the appropriate signal can be issued.

Face detection and recognition is possible with the right camera.   This was done  with the Microsoft Xbox-1 and it is now part of the Apple IPhone X.

There are, of course, limits to how much we want our devices to see and analyze what we are doing.   On the other hand it is clear that advances in automated scene analysis and “understanding” are moving very fast.   Driverless cars are here now and will be commonplace in a few years.   Relatively “smart” robots of various types are under development.   It is essential that we understand how the role of these machines in society can benefit the human condition along the lines of the open letter from many AI experts.

Notes about the MXNet container.

The code is based on a standard example of using MXNet to load a model and invoke it.   To initialize the model, the container first loads the model files into the root file system.  That part is not show here. The files are full-resnet-152-0000.parms (310MB), full-resnet-152-symbols.json (200KB) and full-synset.txt (300KB) . Once loaded into into memory  the full network is well over 2GB and the container requires over 4GB.

Following the load, the model is initialized.

import mxnet as mx
# 1) Load the pretrained model data
with open (' full-synset.txt ','r ') as f:
        synsets = [l.rstrip() for l in f]
sym, arg _params , aux_pa ram s = mx . model .load _checkpoint( 'full-resnet-152' ,0)
# 2) Build a model from the data
mod = mx.mod .Module (symbol =sym , context =mx. gpu ())
mod. bind ( for_training =False, data_shapes=[( 'data ',(1,3,224,224))])
mod. set_params ( arg_params , aux_params )

The function used for the prediction is very standard.  It takes three parameters: the image object, the model and synsnet (the picture labels).   The image is modified to fit the network and then fed to the forward end.  The output is a Numpy array which is sorted and the top five results are returned.

def predict(img, mod, synsets):
    img = cv2.resize(img, (224, 224))
    img = np.swapaxes(img, 0, 2)
    img = np.swapaxes(img, 1, 2)  
    img = img[np.newaxis, :] 
    mod.forward(Batch([mx.nd.array(img)]))
    prob = mod.get_outputs()[0].asnumpy()
    prob = np.squeeze(prob)
    a = np.argsort(prob)[::-1]
    result = []
    for i in a[0:5]:
          result.append( [ prob[i], synsets[i][synsets[i].find(' '):]])
    return result

The container runs as a webservice on port 8050 using the Python “Bottle” package.  When it receives a web POST message to “call_predict” it invokes the call_predict function below.  the image has been passed as a jpeg attachment with is extracted with the aid of the request package.   It is saved in a temporary file and then read by the OpenCV read function.  Unfortunately there was no way to avoid the save followed by read because of limitations to the API.  However we measured the cost of this step and it was less than 1% of the total time of the invocation.

The result of the predict function is a two dimensional array with each row consisting of a probability and the associated label.   The call returns the most likely labels as shown below.

@route('/call_predict', method='POST')
def call_predict():
    t0 = time.time()
    result = ''
    request.files.get('file').save('yyyy.jpg', 'wb')
    image = cv2.cvtColor(cv2.imread('yyyy.jpg'), cv2.COLOR_BGR2RGB)
    t1 = time.time()
    result = predict(image,  mod, synsets)
    t2 = time.time()
    answer = "i think this is a "+result[0][1]+" or it may be a "+result[1][1]
    if result[0][0] < 0.3:          answer  = answer+ ", but i am not sure about this."     if result[0][0] > 0.6:
         answer = "I see a "+result[0][1]+"."
    if result[0][0] > 0.75:
         answer = "This certainly looks like  a "+result[0][1]+"."
    answer = answer + " \n total-call-time="+str(t2-t0)
    return(answer)
run(host='0.0.0.0', port=8050)

The version of the MXNet container used in the ESC experiment replace the Bottle code and call_predict with loop that polls the message queue, pulls a blob from S3 and pushes the result to DynamoDB

In a previous post we described Algorithmia,  a cloud service for discovering, invoking and deploying algorithms.  In this short article we look at Algorithmia as a tool to deploy trained machine learning models.   We used a tool called Gensim to build a model of scientific documents and then create an Algorithmia service that uses the model to predict the topic categories of scientific article.

A Review of Word Vectors and Document Vectors.

This technology has been around for a while now, so this review is more of a history lesson and not a deep technical review.  However, we will give you links to some of the important papers on the subject.

If you want to do document analysis through machine learning, you need a way to represent words in a vector form.  Given a collection of documents you can extract all the words to create a vocabulary.  If the size of the vocabulary is 100,000 words, you can represent each word as a “one-shot” vector in which the i-th word in the vocabulary is a vector of zeros except for a 1 in the i-th position.   Then if your then each document in your collection can be represented as the sum of vectors corresponding the words in that document.   If you have M documents, then the collection is represented by the sparse matrix of size M x 100,000.  Using this “bag of words” representation, there are a variety of traditional techniques such as Latent Sematic Analysis that can be used to extract the similarities between documents.

About five years ago, a team from Google found a much better way to create vectors from words so that words that are used in similar semantic context are nearer to each other as vectors.   Described in the paper by Tomas Mikolov et. all.,  the method, often referred to as Word2Vec, can be considered a map m() of our 100,000 dimension space of word to a dense space of much smaller dimension, say 50, with some remarkable properties.   In particular, there is the now-famous analogy linearity relationships. For example “man is to king as woman is to queen” is expressible (approximately) as

m( king) – m(man) + m(woman)   ≈   m(queen)

There is an excellent set of technical explanations of why Word2Vec work on Quora and we won’t go into them here.  One of the best papers that address this issue is by Golberg and Levy.

Le and Mikolov have shown that the basic methods of Word2Vec generalized to paragraphs, so that we now have a map p() from a corpus of paragraphs to vectors. In other words, given a corpus of documents D of size N, then for any doc d in D, p(d) is a vector of some prespecified length that “encodes” d.  At the risk of greatly oversimplifying, the paragraph vector is a concatenation of a component that is specific to the paragraph’s ID with word vectors sampled from the paragraph.   (As with Word2Vec, there are actually two main versions of this model.  Refer to the Le and Mikolov paper for details.)  It turns out that the function p can be extended to arbitrary documents x so that p(x) is an “inferred” vector in the same space vector space.  We can then use p(x) to find the documents d such that p(d) is nearest to p(x).      If we know how to classify the nearby documents, we can make a guess at the classification of x.  That is what we will do below.

Using Doc2Vec to Build a Document Classifier

Next we will use a version of the Paragraph vectors from Gensim’s Doc2Vec model building tools and show how we can use it to build a simple document classifier.  Gensim is a product of  Radim Řehůřek’s RaRe Technologies.    An excellent tutorial for Gensim is this notebook from RaRe.  To initialize Gensim Doc2vec we do the following.

import gensim
model = gensim.models.doc2vec.Doc2Vec(size=50, min_count=2, iter=55)

This creates a model that, when trained will have vectors of length 50.  The training will use 2 word minimum from each doc for each iteration and there will be 55 iterations.

Next we need to ready a document corpus.   What we will use is 7000 science journal article abstracts from the Cornell University archive ArXiv.org . We have recorded the titles, abstracts and the topic classifications assigned by the authors.   There are several dozen topic categories but we partition them into five major topics: physics, math, computer science, biology and finance.      We have randomly selected 5000 for the training set and we use the remainder plus another 500 from recently posted papers for testing.  We must first convert the text of the abstracts into the format needed by Doc2Vec.  The files are “sciml_train” and “sciml_test”.   The function below preprocesses each of the document abstracts to create the correct corpus.

def read_corpus(fname, tokens_only=False):
        with smart_open.smart_open(fname, encoding="iso-8859-1") as f:
             for i, line in enumerate(f):
                   doc = gensim.utils.simple_preprocess(line)
                   if tokens_only:
                       yield doc
                   else:
                       # For training data, add tags
                       yield gensim.models.doc2vec.TaggedDocument(d, [i])
train_corpus = list(read_corpus("sciml_train"))
test_corpus = list(read_corpus("sciml_test", tokens_only=True))

We next build a vocabulary from the words in the training corpus.   This is a dictionary of all the words together with the counts of the word occurrences.  Once that is done we can train the model.

model.build_vocab(train_corpus)
model.train(train_corpus, total_examples=model.corpus_count, epochs=model.iter)

The training takes about 1 minutes and a simple 4-core server. We can now save the model so that it can be restored for use later with the Python statement model.save(“gensim_model”). We will use this later when building the version we will install in Algorithmia.
The model object contains the 5000 vectors of length 50 that encode our documents. To build our simple classifier we will extract this into an array mar of size 5000 by 50 and normalize each vector to be of unit length. (The normalization will simplify our later computations.)

import Numpy as np
mar = np.zeros((model.docvecs.count, 50))
for i in range(m.count):
    x = np.linalg.norm(model.docvecs[i])
    mar[i] = model.docvecs[i]/x

An interesting thing to do with the mar matrix is to visualize it in 2-d using the t-distributed stochastic neighbor embedding (t-SNE) algorithm. The result is shown in the figure below. The points have been color coded based on topic: 1(dee purple) = “math”, 2(blue gray) = “Physics”, 3(blue green) = “bio”, 4(green) = “finance” and 5(yellow) = “compsci”.

There are two things to note here.   First, the collection is not well balanced in terms of numerical distribution.  About half the collect is physics and there are only a small number of bio and finance papers.   That is the nature of academic science:  lots of physicists publishing papers and not so many quantitative finance or quantitative bio papers in the open literature.   It is interesting to note that the Physics papers divide clearly into two or  three clouds.  (it turns out these separate clouds could be classed as “astrophysics”  and “other physics”.)  Computer science and math have a big overlap and bio has a strong overlap with cs because these are all “quantitative bio” papers.

The classification algorithm is very simple. Our model has a function infer_vector(doc) that will use stochastic methods to interpret the doc into the model vector space. Using that inferred vector we can compute the nearest k documents to it in the model space with the function below.

def find_best(k, abstract):
	preproc = gensim.utils.simple_preprocess(abstract)
	v = model.infer_vector(preproc)
       v0 = v/np.linalg.norm(v)
	norms = []
 	for i in range(5000):
     		norms.append([np.dot(v0,mar[i]), i])
       return norms[0:k]

The dot product of the two normalized vectors is the cosine distance. Because the infer_vector is stochastic in nature, our final version of the classifier calls the find_best ten times and computes an average ranking. (The details are in this notebook. and an Html version.) Selecting one of the more recent abstracts and subjecting it to the classifier gives the result pictured below.

The analysis gives the abstract a score of 80 for computer science and 20 for bio. Note that the title contains the detailed ArXiv category, so we see this is correct, but reading the article it could also be cross listed as bio.
On the other hand, there are many examples that easily confuse the system. For example, the one below is classified as quantitative biology in arXiv, but the system can’t decide if it is math, cs or physics.

In general we can take the highest ranking score for each member of the test set and then compute a confusion matrix.  The result is shown below.  Each row of the table represents the percent of the best guesses from the system for the row label.

One interesting observation here is that in the cases where there is an error in the first guess, the most common mistake was to classify an abstract as mathematics.

Moving the model to Algorithmia

Moving the model to Algorithmia is surprisingly simple.  The first step is to create a data collection in the Algorithmia data cloud.   We created one called “gensim” and it contains the three important files: the gensim model,  topicdict, the dictionary that translates ArXiv topics to our major topics, and the ArXiv topics associated with each of the training documents.   The Algorithmia collection is shown below.  We also loaded the training document titles but they are not necessary.

The main difference between running a trained model in Algorithmia and that of a “normal” algorithm is the part where you load the model from the data container.   The skeleton of the python code now includes a function load_model()which you write and a line that invokes this function as shown below.  Now when your algorithm is loaded into the microservice it first calls the load_model()before invoking the apply(input) function.   For all subsequent invocations of you algorithm while it running in that microservice instance the model is already loaded. (The full source code is here. )

import Algorithmia
import gensim
From gensim.models.doc2vec import Doc2Vec
client = Algorithmia.client()
def load_model():
    file_path = 'data://dbgannon/gensim/gensim_model'
    file_path = client.file(file_path).getFile().name
    model = Doc2Vec.load(file_path)
    # similarly load train_sites and topicdict
    # and create mar by normalizing model data
    return model, mar, topicdict, train_sites

model, mar, topicdict, train_sites = load_model()
def find_best_topic(abstract):
    #body of find_best_topic
def apply(input):
    out = find_best_topic(input)
    return out

Deploying the algorithm follows the same procedure as before.   We add the new algorithm from the Algorithmia portal and clone it.   Assuming the SciDocClassifier.py contains our final version of the source,  we execute the following commands.

git add SciDocClassifier.py
git commit -m "second commit"
git push origin master

Returning to the Algorithmia portal, we can go to the project source editor. From there we need to add the code dependencies. In this case, we wanted exactly the same versions of gensim and Numpy we used in our development environment. As shown below that was easy to specify.

The final version has been published as dbgannon/SciDocClassifer and is available for anyone to use.  Once again, our experience with using Algorithmia’s tools have been easy to use and fun to experiment with. There are many algorithms to try out and a starter free account is all you need.

I was asked to give a talk at Indiana University about cloud computing.   A PDF of the slides are here: What_s New with the Cloud

With a few examples from AWS Lambda, Azure Functions and OpenWhisk.

Cloud computing is going through an interesting evolution.  It has gone from a platform for deploying virtual machines to planet-scale systems with extensive collections of data storage, analysis and machine learning services.   Most recently we have seen the emergence of “cloud native” computing, which in its most basic form involves a design pattern of microservices where big applications are decomposed into hundreds of basic stateless components that run on  clusters managed by tools like Kubernetes, Mesos and Swarm.

Serverless computing is the next step in this evolution.  It addresses the following challenges.  Suppose I have a small computation that I want to run against some database at the end of each month. Or suppose I want to have the equivalent of a computational daemon that wakes up and executes a specific task only when certain conditions arise.  For example, when an event arrives in a particular stream or when a file in a storage container has been modified or when a timer goes off.  How do I automate these tasks without paying for a continuously running server?   Unfortunately, the traditional cloud computing infrastructure model would require me to allocate computing resources such as virtual machines or a microservice cluster and my daemon would be a continuously running process.   While I can scale my cluster of VMs up and down, I can’t scale it to zero without my daemon becoming unresponsive.  I only want to pay for my computing WHEN my computation is running.

This is not a totally new idea.  Paying only for the compute that we use goes back to early timesharing and persists with compute-hour “allocations” on supercomputers today.  And there are cloud services such as Azure Data Lake Analytics, Amazon Kinesis or the AWS API gateway, that charge you only for the computation that you use or data that you move and they do not require you to deploy server infrastructure to use them.

However, there is something deeper going on here and it has to do with triggers and another computational paradigm called “Function-as-a-Service”(FaaS). Unlike the serverless examples above which depend upon me invoking a specific well-defined service, FaaS allows a cloud user to define their own function, and then “register” it with the cloud and specify the exact events that will cause it to wake up and execute. As mentioned above, these event triggers can be tied to changes in state of a storage account or database, events associated with queues or streams of data from IoT devices, web API invocations coming from mobile apps.  Triggers can even be defined by steps in the execution of a workflow.   And, of course, the user only pays when and while the function is executing.

Figure 1.  Function, triggers and output concept. (from markoinsights.com)

There have been two conferences that have focused on the state of serverless computing. The paper “Status of Serverless Computing and Function-as-a-Service (FaaS) in Industry and Research” by Geoffrey Fox, Vatche Ishakian, Vinod Muthusamy and Aleksander Slominski provides an excellent overview of many of the ideas, concepts and questions that surround serverless computing that surfaced at these conferences. In that paper they refer to an IBM tutorial that defines serverless FaaS as
1. short-running, stateless computation
2. event-driven applications
3. scales up and down instantly and automatically
4. based on charge-by-use

Notice that there is a distinction between FaaS and serverless FaaS and it has to do with item 4 above.   A good example of this is Google’s App Engine, which was arguably the first FaaS available from a commercial cloud.   In its current form App Engine can run in one of two modes.  In its standard mode, your applications run in a sandbox and you are charged only when the app is running.  In the “flexible” mode you deploy a container  and then specify the compute infrastructure needed in terms of CPU power, memory, disk and you are charged by the hour.  You could say that App Engine in running Flexible mode is server-lite, but clearly not fully serverless, while standard mode is truly serverless.

What are the serverless FaaS choices?

There are a number of FaaS implementations.  Some of these are used for research while others are commercial products.  The Status report refers to many of these and the slides for the workshop are on-line.   A good example of the research work is OpenLambda from the University of Wisconsin and first introduced dat HotCloud ’16.  Based on this experience the Wisconsin team described Pipsqueak, an experiment to reduce the deployment latencies caused by Python library initializations.   Ryan Chard described Ripple which is an excellent example of distributing event trigger management from the source to the cloud. Ripple has been designed and use for several significant science applications including beamline science (ALS and APS).  Another related technology is if-this-then-that IFTTT that is a service for chaining together other service.

Two other open source projects raise an interesting question about what is behind the curtain of serverless. Funktion  and Fission are both implementations of FaaS on top of Kubernetes. As we discuss serverless computing we must remember that there is a “server” somewhere. The basic infrastructure for serverless computing needs to run somewhere as a persistent service and hence a microservice platform like Kubernetes is a reasonable choice. This relates to the economics of severless computing and we return to that at the end of this report.

The most commonly referenced open source FaaS service is Apache OpenWhisk which was developed by IBM and is available on their Bluemix cloud as a service. The other commercial services include Google Function, Microsoft Azure Functions and Amazon Lambda. At the end of this article we will show some very simple examples of using some of these systems.

When can FaaS replace a mesh of microservices for building an app?

The Status paper also makes several other important observations. For example, they note that while serverless is great for the triggered example described above, it is not good for long-running or statefull applications like databases, deep learning training, heavy stream analytics, Spark or Hadoop analysis and video streaming. In fact, many large-scale cloud-native applications that have thousands of concurrent users require continuous use of massive networks of microservices. This will will not be based on serverless FaaS. However, there may be many cases where a user-facing application running in the cloud could be easily implemented with serverless FaaS rather than as a big microserivce deployment. What we do not know is where the cross-over point from a serverless FaaS implementation of an app to a full kubernetes based massive microservice deployment lies. This relates to the economic of FaaS (discussed briefly at the end of this article). The Cloud Native Computing Foundation has a working group on serverless computing that is addressing this topic.

There is one interesting example of using serverless FaaS to do massively parallel computing and it called pywren. The lovely paper “Occupy the Cloud: Distributed Computing for the 99%”  by Eric Jonas, Qifan Pu, Shivaram Venkataraman, Ion Stoica and Benjamin Recht describes the concepts in pywren which allow it to scale computation to thousands of concurrent function invocations achieving 40teraflops of compute performance. Pywren uses AWS Lambda in a very clever way: it serializes the computational function which are then passed to a lambda function to execute. We will return to Pywren in another post.

Computing at the Edge?

Perhaps the hottest topic in the general area of cloud computing is when the computing spills out of the cloud to the edge of the network. There are many reasons for this but most boil down to latency and bandwidth. A good example is the use cases that motivate the Ripple system described above. In order to generate a stream of events from sensors at the edge of the network, one needs very light weight computing that can monitor them and generate the events. In many cases it is necessary to preprocess the data in order to send the message to the cloud where a function will be invoked to respond. In some cases, the response must return to the source sooner than a remote invocation can respond because of the latencies involved. The computing at the edge may need to execute the function there and the actual communication with the cloud may be just a final log event or a trigger for some follow-up action.

Another possibility is that the functions can migrate to the places where they are needed. When you deploy the computing at the edge you also deploy a list of functions that that must be invoked. When the edge process starts up it could cache some of the function it needs locally. We expect there will be many variations on these ideas in the future.

A Tiny FaaS Tutorial

By their very nature FaaS systems are very simple to use: you write short, stateless function and tie them to triggers. We will take a quick look at three of these: AWS Lambda, Microsoft Function and IBM Bluemix OpenWhisk.

AWS Lambda Functions

Amazon was the first to introduce serverless functions as a service and it is very well integrated into their ecosystem of other services. Called Lambda functions, they can be easily created in a number of standard programming languages. Lambda functions can be associated with a variety of trigger events including changes to the state of a storage account, web service invocations, stream events and even workflow events.

We will illustrate Lambda with a simple example of a function that responds to Kinesis Stream events and for each event it adds an item to a dynamoDB table. Here is the python code for the function that accomplishes this task.

from __future__ import print_function
#import json
import base64
import boto3
def lambda_handler(event, context):
    dyndb = boto3.resource('dynamodb', region_name='us-west-2')
    table = dyndb.Table("lambdaTable")
    
    for record in event['Records']:
       #Kinesis data is base64 encoded so decode here
       payload=base64.b64decode(record["kinesis"]["data"])
       x = eval(str(payload))
       metadata_item = {'PartitionKey': x['PartitionKey'], 
                        'RowKey': x['RowKey'], 'title': x['text']}
       table.put_item(Item=metadata_item)

There are several critical items that are not explicit here.   We need to invoke the AWS Identity and Access Management (IAM) system to delegate some pemissions to our function and we will need to grab copies of the Amazon Resource Names (ARNs) for this and other objects. First we create a IAM role that will allow access to Kineses streams and DynamoDB. Using the IAM portal we created a role called:  “lambda-kinesis-execution-role” and attached two policies ‘AmazonDynamoDBFullAccess” and “AWSLambdaKinesisExecutionRole”. We then made a copy of the Role ARN.

The next step is to install this function into the lambda system. To do that we put the function above into a file called “ProcessKinesisRecords.py” and then zipped it. We then uploaded the zipped file to S3 (in the us-west-2) in the bucket “dbglambda”. With that we can create our function with the boto3 call from our laptop as:

lambdaclient = boto3.client('lambda')

response = lambdaclient.create_function(
    FunctionName='lambdahandler',
    Runtime='python2.7',
    Role='arn:aws:iam::066xx534:role/lambda-kinesis-execution-role',
    Handler='ProcessKinesisRecords.lambda_handler',
    Code={
        'S3Bucket': 'dbglambda',
        'S3Key': 'ProcessKinesisRecords.zip',
    }
    )

The final item we need is the link between this function and the Kinesis Stream service.
To do that we went to the portal for Kinesis and created a stream called “mylambdastream” and grabbed its ARN. Creating the binding is accomplished with the following.

response = lambdaclient.create_event_source_mapping(
    EventSourceArn=
        'arn:aws:kinesis:us-west-2:06xxx450734:stream/mylambdastream',
    FunctionName='lambdahandler',
    BatchSize=123,
    StartingPosition='TRIM_HORIZON',
)

We can verify the properties of the function we have created by looking at the AWS Lambda portal pages. As shown below, we can verify that our lambdahandler function does indeed have our stream as its trigger.

Figure 2. AWS Lambda Function dashboard showing our trigger.

Finally we can invoke it by pushing an event to the kinesis stream. This is shown below.

client = boto3.client('kinesis')
record = "{'PartitionKey':'part2', 'RowKey': '444', 'text':'good day'}"
resp = client.put_record(StreamName='mylambdastream',
    Data=bytearray(record),
    PartitionKey='a')

Checking the DynamoDB portal will verify that the function has picked the message from Kinesis and deposited it in the database. The full details are in the notebook “simple-kinesis-lambda-sender.ipynb”.

Azure Functions

The Azure function portal has a large number of basic templates we can use to build our function as shown below.

Figure 3.  Azure Function Template samples

We have selected one of the Python examples that creates a simple web service. Bringing this up on the portal we see the code below.

import os
import json

postreqdata = json.loads(open(os.environ['req']).read())
response = open(os.environ['res'], 'w')
response.write("hello world to "+postreqdata['name'])
response.close()

To get this running, we now must go to the “integrate” page to provide one more detail: we set the authorization level to “anonymous” as shown below.

Figure 4.  Setting the authorization level for the trigger to “anonymous”.

The client is also simple and shown below.

import os
import json
import requests
import json

data = {}
data['name'] = 'goober'
json_data = json.dumps(data)
r = requests.post("https://pylook.azurewebsites.net/api/pylook", 
                   data=json_data)
print(r.status_code, r.reason)
print r.text

Using Azure Functions to pull data from a queue.

We now give another version of the function that reads messages from a queue and puts them in a table. There is no Python template for this one yet, so we will use JavaScript. To use template this we must first create a new storage account or use an existing one. Go to the storage account page and you will see

Figure 5. Creating a new table and queue in a storage account.

We click on Table and create a new table and remember its name. Then we go to Queues and create a new one and remember its name. Looking in the storage explorer should show for these items. Clicking on the table name should give us a picture of an empty table.

Go back to the portal main page and click + and look for “Function App” and click create. There is a table to fill in like the one below.

Figure 6.  Creating a new function.

We give it a name and allow it to create a new resource group. For the storage we want to use the dropdown and look for the storage account name. (It is important that the storage account is in the same location as the function. We click create and wait for the function to appear on the function portal page. You should see it in your resource groups. Follow that link and you will see that it is running.

Go to the functions tab and hit “+”. It will ask you to pick one of the templates. At the top where it says “language” select Javascript and pick the one that is called QueueTrigger. This will load a basic template for the function. Now edit the template so that it looks like the following example.

The main difference between this and the template is that we have added an output table and instructions to push three items into the table. The function is assuming that the items in the queue are of the form

{‘PartitionKey’: ‘part1’, ‘RowKey’: ’73’, ‘content’: ‘some data ‘}

Next we need to tie the queue to our queue and the table to our table. So click on “Integrate” on the left. You need to fill in the form so that it ties it to your stuff as illustrated below.

Figure 8.  The association between the storage account queue and table service with the variable in our function.  Here we have highlighted the “show value” to verify that it has the right storage account.

You should see your storage account in the dropdown menu. Select it. And they add the Table name. You need to do the same for the AzureQueueStorage. Once this is done and your function is saved and the system is running your function should be instantiated and invoked as soon as you send it queue items. For that we have a simple python script in a Jupyter notebook. You can get from  https://SciEngCloud.github.io/py-functions-queue-driver.ipynb.
You will need to fill in your account key for the storage account, but then you should be able to step through the rest. The notebook runs a few tests and then sends 20 items to the queue. Using the Azure storage explorer we see the results in the table as shown below.

Figure 9.   The view of the table after running the Jupyter Notebook

OpenWhisk and IBM Bluemix

OpenWhisk is the open source serverless function system developed by IBM that is also supported as a commercial service on their Bluemix cloud.  Like the others it is reasonably easy to use from their command-line tools, but the Bluemix portal provides an even easier solution.   Creating a function is as easy as selecting the runtime and downloading a template.  Figure xx below illustrates a simple function derived from template for a Python3 function.

Figure 10.  An OpenWhisk function that decodes a dictionary input and returns a dictionary.

Notice that the parameter to the function is a python dictionary (all Openwhisk messages are actually Json objects, which for Python are rendered as dictionaries).  While this template can be run as a web service in its present form, it is also possible to connect it to an external trigger.  To illustrate this we connected it to a trigger  that monitors activity in a Github account.  Using the portal it was easy to create the trigger (called “mygithubtrigger” and bind it to push and delete actions on a repository called dbgannon/whisk.   All this required was an access token which was easy available by logging in to the GitHub portal.   In the case of GitHub triggers the event returns a massive Json object that is a complete description of the repository.   In the action “myaction” we go two levels down in the dictionary and extract the official repository description and make that the response of the action to the event.

When a trigger fires you need a rule which bind the firing to an action.   We bound it to the trivial “myaction” example above.  The view from the portal of the rule is below

Figure 11.   The rule view of our association of the trigger to the function

We next added a new file to the repository.  This activated the trigger and the rule invoked the action.   The portal has a nice monitor facility and the image is below.

Figure 12.  This is the “Monitor” view after the trigger has fired.

Finally, drilling down on the “myaction” log event we see the description of the GitHub repository we created.

Figure 13.  The output of the “myaction” function after a file was added to the GitHub repository.

Finally

These examples above are all very trivial.   The next thing to explore is how functions can be composed into workflows.   Each of the three systems has its own way of doing that and if we have time later we will show some examples of this capability.   We have also not disussed performance or cost which is greatly dependent on the rate at which your triggers are firing and the amount of work in each function execution.

The economics of serverless computing are also very interesting.  As we pointed out  earlier, for a cloud provider to offer a FaaS serverless capability it must be supported by actual infrastructure.   How can the provider afford this if it is only charging by the second?  There are two possible answers.   First if your FaaS is serving very large numbers of function per second then it will certainly pay for itself.   But there is another consideration.   Many of the big cloud providers are running vast microservice frameworks that support most, if not all of their big internal and publiclly available applications.  Running  a FaaS system on top of that is as easy as running any other microservice based application.  One only needs to set the cost per second of function evaluation high enough to cover the cost of that share of the underlying platform the FaaS system is using.   For Amazon Lambda “The price depends on the amount of memory you allocate to your function. You are charged $0.00001667 for every GB-second used.”  This amounts to $0.06 per GB hour.  So for a function that take 8 GB of memory to execute that is $.48 per hour, which is 4.8 times greater than the cost of an 8GB m4.large ec2 instance.   Consequently a heavily used FaaS system is a financial win and a lightly used one may have little impact on your infrastructure.

Of course, single executions of a FaaS function are limited to five minutes, so you would need 12 5 minute concurrent executions to reach an hour of execution time. Furthermore AWS give you 1 million invocations (at 100ms) each month for free or 400,000 GB-seconds per month free.  That is 111 GB hours per month for free.   This is not a bad deal and it would seem to indicate that Lambda is not a big drain on their infrastructure yet.

Acknowledgement:  Thanks go to Ryan Chard for helpful comments and suggestions on a draft of this post.