Stream processing is experiencing exponential growth with businesses and services relying heavily on real-time analytics, inferencing, monitoring, and more. Reliable, cost-effective streaming at scale is paramount, but auto-scaling has hit cost-efficiency limits with CPUs. This talk will be about how NVIDIA is leveraging Dask to GPU-accelerate big data stream processing at scale in production.
Desperate to solve the problem of how we, at NVIDIA, could dramatically scale the processing of streaming data while reducing infrastructure costs, we set out to develop cuStreamz - the world’s first GPU-accelerated streaming library.
This talk will discuss how we built and are leveraging cuStreamz, our end-to-end GPU-accelerated streaming library - whose core comprises RAPIDS, Streamz, and of course, Dask.
cuStreamz boosts end-to-end streaming throughput as well as lowers the total infrastructure cost for running complex streaming pipelines at scale. Moreover, folks can start building streaming pipelines using simple Python API!
cuStreamz offers out-of-the-box GPU-accelerated data readers & writers for commonly used data formats like JSON, Parquet, CSV, etc. And as for data ingestion, it has a custom GPU-accelerated Kafka data source, which has proven to read data at least 2x times faster from Kafka than the widely used Confluent Kafka Python library.
cuStreamz is also Kubernetes-ready, i.e., anyone can launch a cuStreamz compute cluster on a bunch of GPUs anywhere in the world, and scale up and down, with really minimal effort.
We will also talk about some very interesting (and complex) Cloud Gaming streaming use cases at NVIDIA which are currently running in production on cuStreamz.
Come learn how you can run streaming at scale on GPUs with really simple Python API and how we are getting ~3x+ cost savings compared to the state-of-the-art Spark Streaming while running streaming pipelines in production.