HHMI’s Janelia Research Campus in Ashburn, Virginia has an integrated team of lab scientists and tool-builders who pursue a small number of scientific questions with potential for transformative impact. To drive science forward, we share our methods, results, and tools with the scientific community.
Our neuroscience research application involves image searches that are computationally intensive but have unpredictable and sporadic usage patterns. The conventional on-premises approach is to purchase a powerful and expensive workstation, install and configure specialized software, and download the entire dataset to local storage. With 16 cores, a typical search of 50,000 images takes 30 seconds. A serverless architecture using AWS Lambda allows us to do this job in seconds for a few cents per search, and is capable of scaling to larger datasets.
Parallel Computation in Neuroscience Research
Basic research in neuroscience is often conducted on fruit flies. This is because their brains are small enough to study in a meaningful way with current tools, but complex enough to produce sophisticated behaviors. Conducting such research nonetheless requires an immense amount of data and computational power. Janelia Research Campus developed the NeuronBridge tool on AWS to accelerate scientific discovery by scaling computation in the cloud.
Figure 1: A “mask image” (on the left) is compared to many different fly brains (on the right) to find matching neurons. (Janella Research Campus)
The fruit fly (Drosophila melanogaster) has about 100,000 neurons and its brain is highly stereotyped. This means that the brain of one fruit fly is similar to the next one. Using electron microscopy (EM), the FlyEM project has reconstructed a wiring diagram of a fruit fly brain. This connectome includes the structure of the neurons and the connectivity between them. But EM is only half of the picture. Once scientists know the structure and connectivity, they must perform experiments to find what purpose the neurons serve.
Flies can be genetically modified to reproducibly express a fluorescent protein in certain neurons, causing those neurons to glow under a light microscope (LM). By iterating through many modifications, the FlyLight project has created a vast genetic driver library. This allows scientists to target individual neurons for experiments. For example, blocking a particular neuron of a fly from functioning, and then observing its behavior, allows a scientist to understand the function of that neuron. Through the course of many such experiments, scientists are currently uncovering the function of entire neuronal circuits.
We developed NeuronBridge, a tool available for use by neuroscience researchers around the world, to bridge the gap between the EM and LM data. Scientists can start with EM structure and find matching fly lines in LM. Or they may start with a fly line and find the corresponding neuronal circuits in the EM connectome.
Both EM and LM produce petabytes of 3D images. Image processing and machine learning algorithms are then applied to discern neuron structure. We also developed a computational shortcut called color depth MIP to represent depth as color. This technique compresses large 3D image stacks into smaller 2D images that can be searched efficiently.
Image search is an embarrassingly parallel problem ideally suited to parallelization with simple functions. In a typical search, the scientist will create a “mask image,” which is a color depth image featuring only the neuron they want to find. The search algorithm must then compare this image to hundreds of thousands of other images. The paradigm of launching many short-lived cloud workers, termed burst-parallel compute, was originally suggested by a group at UCSD. To scale NeuronBridge, we decided to build a serverless AWS-native implementation of burst-parallel image search.
Our main reason for using a serverless approach was that our usage patterns are unpredictable and sporadic. The total number of researchers who are likely to use our tool is not large, and only a small fraction of them will need the tool at any given time. Furthermore, our tool could go unused for weeks at a time, only to get a flood of requests after a new dataset is published. A serverless architecture allows us to cope with this unpredictable load. We can keep costs low by only paying for the compute time we actually use.
One challenge of implementing a burst-parallel architecture is that each Lambda invocation requires a network call, with the ensuing network latency. Spawning several thousands of functions from a single manager function can take many seconds. The trick to minimizing this latency is to parallelize these calls by recursively spawning concurrent managers in a tree structure. Each leaf in this tree spawns a set of worker functions to do the work of searching the imagery. Each worker reads a small batch of images from Amazon Simple Storage Service (S3). They are then compared to the mask image, and the intermediate results are written to Amazon DynamoDB, a serverless NoSQL database.
Figure 2: Serverless architecture for burst-parallel search
Search state is monitored by an AWS Step Functions state machine, which checks DynamoDB once per second. When all the results are ready, the Step Functions state machine runs another Lambda function to combine and sort the results. The state machine addresses error conditions and timeouts, and updates the browser when the asynchronous search is complete. We opted to use AWS AppSync to notify a React web client, providing an interactive user experience while remaining entirely serverless.
As we scaled to 3,000 concurrent Lambda functions reading from our data bucket, we reached Amazon S3’s limit of 5,500 GETs per second per prefix. The fix was to create numbered prefix folders and then randomize our key list. Each worker could then search a random list of images across many prefixes. This change distributed the load from our highly parallel functions across a number of S3 shards, and allowed us to run with much higher parallelism.
We also addressed cold-start latency. Infrequently used Lambda functions take longer to start than recently used ones, and our unpredictable usage patterns meant that we were experiencing many cold starts. In our Java functions, we found that most of the cold-start time was attributed to JVM initialization and class loading. Although many mitigations for this exist, our worker logic was small enough that rewriting the code to use Node.js was the obvious choice. This immediately yielded a huge improvement, reducing cold starts from 8-10 seconds down to 200 ms.
With all of these insights, we developed a general-purpose parallel computation framework called burst-compute. This AWS-native framework runs as a serverless application to implement this architecture. It allows you to massively scale your own custom worker functions and combiner functions. We used this new framework to implement our image search.
The burst-parallel architecture is a powerful new computation paradigm for scientific computing. It takes advantage of the enormous scale and technical innovation of the AWS Cloud to provide near-interactive on-demand compute without expensive hardware maintenance costs. As scientific computing capability matures for the cloud, we expect this kind of large-scale parallel computation to continue becoming more accessible. In the future, the cloud could open doors to entirely new types of scientific applications, visualizations, and analysis tools.
We would like to express our thanks to AWS Solutions Architects Scott Glasser and Ray Chang, for their assistance with design and prototyping, and to Geoffrey Meissner for reviewing drafts of this write-up.
All of the application code described in this article is open source and licensed for reuse:
The data and imagery are shared publicly on the Registry of Open Data on AWS.