Visiting Scientist at A*CRC (A*STAR Computational Resource Centre)
Visiting Professor at the Computer Science Department of National University of Singapore (NUS)
Former Director at Intel Labs and former Chief Product Architect at AMD (USA)
“Irreproducibility: The Causes and The Cure”
John L. Gustafson, A*CRC, NUS
The use of cloud computing resources is increasingly exposing users to the fact that results of large-scale computations are not identical across systems, or even on the same system running the same data set. Some have started using the term “floating point drift” to describe, say, how a physical simulation of the climate will produce two different results on two different HPC systems. Such inconsistency creates a lack of confidence that either simulation is correct, as well it should.
Aside from parallel programming errors that create race conditions or assume memory coherence that is not provided by the system, the causes of inconsistencies stem from key decisions in the IEEE 754 floating point standard. These include the decision that transcendental functions need not be properly rounded for all input arguments, and the encouragement of the use of “guard bits” to reduce rounding errors in a way completely invisible to the programmer. We show the sources of errors and then show a new paradigm that eliminates the irreproducibility problem, rendering calculations with real numbers as bit-for-bit reproducible as pure integer calculations. Furthermore, enforcing the paradigm actually can increase the speed of calculations instead of slowing them down.
Wednesday 7 February 2018, 9:25 am – 10:10 am
Dr. John L. Gustafson is an applied physicist and mathematician who is a Visiting Scientist at A*CRC and Professor at NUS, where he is directing efforts in Next-Generation Arithmetic.
He is a former Director at Intel Labs and former Chief Product Architect at AMD. A pioneer in high-performance computing, he first demonstrated scalable massively parallel performance on real applications in 1988. This became known as Gustafson’s Law, for which he won the inaugural ACM Gordon Bell Prize. He is also a recipient of the IEEE Computer Society’s Golden Core Award.
“The New Stack”
Wednesday 7th February 2018, 11:15 am – 12:15 pm
“Artificial Intelligence, Blockchain, Deep Learning, IoT… what’s behind the hype?”
Friday 9th February 2018, 11:15 am – 12:15 pm
“Meltdown & Spectre: and now what?”
Friday 9th February 2018, 3:30 pm – 4:25 pm
— — — —
- Wikipedia – Linkedin
- What next for supercomputing and problem solving? (Interview with Radio New Zealand at Multicore World 2016 – 17 February 2016)
- CoNGA (Conference for Next Generation Arithmetic) March 2018, Singapore.
Featured Image (top): Keynote at Multicore World 2014
Panel: Multicore World 2017