ICPP 2014
Keynotes Summary
09/24
Data Centric Systems: The Next Pa radigm in Computing
Speaker: Dr. Tilak Agerwala
◦ Vice President, Data Centric Systems
◦ IBM T.J. Watson Research Center
2014/09/09
Data-Centric System
HPE (High Performance Environment)
◦ = HPC + HPA
Mixed compute capabilities require d.
◦ Heterogeneity is important.
IBM Data-Centric Design Principle s
Minimize data motion
Enable compute in all levels of th e systems hierarchy
Modularity
Application design
Leverage OpenPOWER
High Performance Computing - Futu re Directions
Speaker: Prof. Jack Dongarra
◦ University of Tennessee, Knoxville
2014/09/10
Top500 Factoids
There are 37 systems > Pflop /s (up 6 from November).
About 90% of all the systems on the Top500 list are inte grated by U.S. vendors, including 65 of the 76 Chinese s upercomputers.
HP has 182 systems on this list, or more than 36%, follo wed by IBM with 176, or 35%. Cray has 50 or 10%, SGI at 19 systems, and Dell at 8 systems.
Intel processors largest share, 87% followed by AMD, 6%.
For the first time, < 50% of Top500 are in the U.S. -- j ust 233 of the systems are U.S.-based, China #2 w/76.
IBM’s BlueGene/Q is still the most popular system in th e TOP10 with four entries.
Infiniband found in 221 systems, GigE in 202, 10-GigE in 75.
Issue: Memory Transfer
Communication bounded operation
◦ Real performance < peak performance
◦ “Its all about data movement”
◦ Ex:
Take two double precision vectors x and y of size n=375,000.
Time to move the vectors from memory to ca che:
(6MBytes) / (25.6GBytes/sec) = 0.23ms
Time to perform computation of DOT:
(2n flop) / (56Gflop/sec) = 0.01ms
