CUDA Spotlight: GPU-Accelerated Large-Scale Analytics
This week's spotlight is on Dr. Ren Wu, Senior Research Scientist at HP Labs in Palo Alto, Calif., and Principal Investigator of the CUDA Research Center at HP Labs. Dr. Wu works in the area of large-scale analytics and business intelligence.
NVIDIA: Ren, tell us about what you are working on.
Ren: It is no secret that large-scale analytics pose challenges and offer opportunities at same time. The exponential increase in the size of datasets, the desire for real-time or near real-time responses, and the need to use more sophisticated algorithms to gain insights from the data, have made big data analytics a critical research area. It promises to offer tremendous competitive advantage for those who excel in it. Here at HP Labs we have a team dedicated to answering this challenge. I am a member of this team, with a special interest in figuring out what role the GPU will play in enabling next-generation real-time business intelligence.
NVIDIA: What are some potential applications of GPU-accelerated large-scale data analytics?
Ren: There are many different and exciting scenarios, including trend analysis of social media, large-scale predictive models, automatic data correlation, and statistical analysis in sensor data streams. And the GPU is not limited to accelerating large-scale analytics only; it can be used for accelerating analytics on any scale. I expect to see many more GPU-accelerated cloud services in the near future.
NVIDIA: How does GPU computing play a role in your work?
Ren: In today’s world, we’ve been pretty successful in storing and managing large volumes of data, even at petabytes and beyond. However, to make sense out of this data and to do it in a timely fashion, we are not quite there yet. At HP Labs, we are investigating the following: where and how these analytics should be run, what kind of parallel algorithms we should use, how to partition the data and reduce data movement, and how the algorithms can be accelerated. We are looking at all of these questions, and I am focused on the GPU/CUDA acceleration. I believe it is a very important piece of this puzzle. So far, it has been very promising.
NVIDIA: What kind of advantages have you achieved with CUDA?
Ren: We’ve looked at various core primitives in analytics. Per chip comparison, typically we’ve achieved about a 5-20X performance advantage by using CUDA GPUs over a pure CPU approach. For a typical workstation with two CPUs and two Tesla boards, this ratio is also pretty much maintained. Same goes for the servers, even though it can vary more, depending on the server’s configuration.
Personally, I think that the CUDA programming model is a very nice framework – well balanced on abstraction and expressing power, enough control for algorithm designers, and supported by hardware with exceptional performance (compared to other alternatives).
In particular, I like the fact that the algorithm designer can manually manage cache (shared memory to global memory, and global memory to PC memory). For me, one of the key requirements for high performance computing on a many-core architecture is the ability to optimize against the memory hierarchy.
NVIDIA: How did you become interested in this area?
Ren: I am the kind of person who likes to push performance to its extreme. I was intrigued by the potential of GPGPU even during the pre-CUDA era. I did some extensive research back then to see if I could use it to speed up various machine learning algorithms. While I was impressed by the potential performance gain, I concluded that it was not ready for prime time.
Of course, the release of CUDA changed everything. When I first saw CUDA, I knew instantly that this could be a game changer. That was when we started to look at GPU computing seriously.
NVIDIA: As computing becomes faster, what will we be able to do in the future that we are not able to do today?
Ren: Faster computing, combined with availability of large data sets, will enable us to build predictive models which would be too slow or too costly to build today. Running these powerful models against real-time streaming data will allow us to monitor areas such as our environment, health, and financial systems, in a way not possible before – resulting in new insights. Many breakthroughs will follow. I cannot wait to see that happen!
Dr. Ren Wu’s bio
Dr. Ren Wu is a Senior Research Scientist at HP Labs, Palo Alto. His research interests include data-intensive high-performance computing, massively parallel algorithms and computational intelligence. In recent years he has been focusing on GPU acceleration of large-scale analytics, and is well known for his work on GPU-accelerated clustering algorithms. He is also the PI of the CUDA Research Center at HP Labs.