It's the end of the world as we know it

Transcription

1 It's the end of the world as we know it Simon McIntosh-Smith University of Bristol HPC Research Group 1

2 Background Graduated as Valedictorian in Computer Science from Cardiff University in 1991 Joined Inmos to work for David May as a microprocessor architect Moved to Pixelfusion in 1999 a high- tech start- up designing the first GPGPU, a many- core general purpose graphics processor Co- founded ClearSpeed in 2002 as Director of Architecture and ApplicaMons Joined the CS department at the University of Bristol in April 2009 to focus on High Performance CompuMng. Member of OpenCL standards body, Khronos, contributed to design of Archer 2

3 Business as usual? We've all got used to new machines being relatively simple evolutions of our previous machines This will no longer be true from 2016 onwards 3

4 4

5 What's changing? Mainstream multi-core CPUs will continue to evolve, but more slowly IVB à Haswell à Broadwell CPUs 12 core à 18 core à >20 core To retain the levels of performance increase we have historically enjoyed, we will have no choice but to adopt radically different architectures 5

6 What are the options? Many-core CPUs: Intel Xeon Phi Knights Landing (KNL) launching late 2015 Large KNL machines going into US national labs O(70) cores, 512-bit wide vectors, HBM Other many-core CPUs expected to emerge Based on ARM architecture Multiple vendors AMD, Broadcom, Cavium, AMCC, 6

7 What are the options? GPUs: Nvidia Pascal Tightly couples with IBM Power9 AMD Actually have more memory bandwidth and FLOPs than Nvidia Interesting focus on tight CPU/GPU integration ("fat APUs") 7

8 What other big changes are coming? Deeper memory hierarchies Stacking (HBM), non volatile memories etc. Integrated interconnects E.g. Intel Omni-Path fabric FPGAs? Now supporting OpenCL 8

9 Long- term fundamental trends Relative improvement We need to design codes for here! We design codes for here Microprocessor performance ~55% per annum Memory capacity ~49% per annum (and slowing down?) Memory bandwidth ~30% per annum (and slowing down?) Memory latency <<30% per annum Time

10 Will this affect me? 10

11 Data from Top500 5 of top 10 leadership class machines already rely on Xeon Phi or GPUs Holdouts are mostly based on IBM's BlueGene/Q, but this is end-of-lifeing From 2016 onwards, expect almost all of the Top 10 machines world-wide to rely on these radically different architectures Performance chasm will then open up between "haves" and "have nots" 11

12 Where is Archer in Top500? Entered at #19, now at #25 6 more Phi/GPU machines between Archer and the Top10 Ratio has been rising steadily I.e. son of Archer likely to be surrounded by many-core machines in the Top

13 Landscape of HPC in the UK Now on to some better news! The UK has a vibrant and world-class HPC community Diverse range of provision: National: Archer, DiRAC, STFC etc. Regional / thematic: SES, N8, HPC Wales etc University-level facilities 30 UK machines in Nov 2014 Top500 (16 academic) 13

14 Landscape of HPC in the UK Storage facilities: Research Data Facility (~14PB disk + 19PB tape) Jasmine (NERC) ~4,000 cores, approaching 10 PB disk Support for developing scientific codes: Embedded CSE support Sustainable Software Institute (SSI) NAG Software development fellowships 14

15 Summary A "business as usual" approach to scientific software development will result in being left in the slow lane Developers are faced with the challenging issue of developing performant code on increasingly complex architectures We will all need a lot of support from the RCs, HPC services ecosystem et al! 15