Last week, I attended the annual IEEE High Performance Extreme Computing Conference (formerly called High Performance Embedded Computing) in Waltham, MA. I had the privilege of presenting my paper on distributed database performance, and I got some great comments and questions.
Here are the key themes that came across in many of the talks and keynotes:
1. Hadoop MapReduce is entering the “trough of disillusionment”
The MapReduce programming pattern is inadequate for all but the simplest of analytics. On top of that, the Hadoop implementation of this classic model of parallelism is bogged down by a weak scheduler and inefficient middleware. Looking ahead, it can serve “embarrassingly parallel” applications, but future versions need to address some of the performance problems.
2. The next generation of intelligent analytics rely on sparse computation
In years past, the HPEC community was laser-focused on signal processing and accelerators such as FPGAs and GPUs. The engineer’s goal was to squeeze every last FLOP out of a computing system. This year, about 10 talks dealt with applications of sparse matrices and data structures. This is a major shift.
3. Tomorrow’s chips are going to look a lot different
Intel’s latest commercial chips are 22nm. Moore’s Law will get transistors down to 5-10nm. After that? This community has to innovate. Talks from Intel, Texas Instruments, MIT, Carnegie Mellon and others mentioned tricks like 3D chip stacking, making memory smarter, and other novel architectures as ways to cram more transistors on a chip, move data faster, and accommodate future applications that look nothing like the dense computation for which today’s systems were optimized.
Reach out on Twitter or leave a reply below if you noticed other themes!