mtaap’07 keynote - pnnlhpc.pnl.gov/mtaap/mtaap07/mtaap_files/keynote.pdf · 2014. 2. 7. ·...
TRANSCRIPT
![Page 1: MTAAP’07 Keynote - PNNLhpc.pnl.gov/mtaap/mtaap07/mtaap_files/keynote.pdf · 2014. 2. 7. · MTAAP’07 Keynote Michael Merrill. Outline What’s important... applications ... Processor](https://reader034.vdocuments.site/reader034/viewer/2022051912/600357b9e26afc0b2d54480b/html5/thumbnails/1.jpg)
MTAAP’07 KeynoteMichael Merrill
![Page 2: MTAAP’07 Keynote - PNNLhpc.pnl.gov/mtaap/mtaap07/mtaap_files/keynote.pdf · 2014. 2. 7. · MTAAP’07 Keynote Michael Merrill. Outline What’s important... applications ... Processor](https://reader034.vdocuments.site/reader034/viewer/2022051912/600357b9e26afc0b2d54480b/html5/thumbnails/2.jpg)
Outline
What’s important... applicationsMaking sense... of all this stuffWhat’s necessary... I thinkWhat’s possible... maybe
![Page 3: MTAAP’07 Keynote - PNNLhpc.pnl.gov/mtaap/mtaap07/mtaap_files/keynote.pdf · 2014. 2. 7. · MTAAP’07 Keynote Michael Merrill. Outline What’s important... applications ... Processor](https://reader034.vdocuments.site/reader034/viewer/2022051912/600357b9e26afc0b2d54480b/html5/thumbnails/3.jpg)
What’s important... applications
Data Structures are important... more support for linked data structures
Ignored algorithm areas are coming back to bite us
Sparse methods on unstructured data
Adaptive methods are better aligned with nature but not with current architecture
Helping humans deal with information overload
![Page 4: MTAAP’07 Keynote - PNNLhpc.pnl.gov/mtaap/mtaap07/mtaap_files/keynote.pdf · 2014. 2. 7. · MTAAP’07 Keynote Michael Merrill. Outline What’s important... applications ... Processor](https://reader034.vdocuments.site/reader034/viewer/2022051912/600357b9e26afc0b2d54480b/html5/thumbnails/4.jpg)
Making sense... of all this stuff
![Page 5: MTAAP’07 Keynote - PNNLhpc.pnl.gov/mtaap/mtaap07/mtaap_files/keynote.pdf · 2014. 2. 7. · MTAAP’07 Keynote Michael Merrill. Outline What’s important... applications ... Processor](https://reader034.vdocuments.site/reader034/viewer/2022051912/600357b9e26afc0b2d54480b/html5/thumbnails/5.jpg)
Spectrum
Hardware contexts per set of functional units
all contexts to one set of
functional units
one context to one set of
functional units
MTA/XMT UltraSparc T1 Cyclops64
![Page 6: MTAAP’07 Keynote - PNNLhpc.pnl.gov/mtaap/mtaap07/mtaap_files/keynote.pdf · 2014. 2. 7. · MTAAP’07 Keynote Michael Merrill. Outline What’s important... applications ... Processor](https://reader034.vdocuments.site/reader034/viewer/2022051912/600357b9e26afc0b2d54480b/html5/thumbnails/6.jpg)
Stuff
VirtualizationHow much is enough?Fault tolerance
How much baggage does a context have? Probably affects virtualizationSynchronization
![Page 7: MTAAP’07 Keynote - PNNLhpc.pnl.gov/mtaap/mtaap07/mtaap_files/keynote.pdf · 2014. 2. 7. · MTAAP’07 Keynote Michael Merrill. Outline What’s important... applications ... Processor](https://reader034.vdocuments.site/reader034/viewer/2022051912/600357b9e26afc0b2d54480b/html5/thumbnails/7.jpg)
More Stuff
Explicit memory hierarchy?I-cache! Don’t make the programmer worry about code size!?!Commercial use vs. scientific use... vs. something else
![Page 8: MTAAP’07 Keynote - PNNLhpc.pnl.gov/mtaap/mtaap07/mtaap_files/keynote.pdf · 2014. 2. 7. · MTAAP’07 Keynote Michael Merrill. Outline What’s important... applications ... Processor](https://reader034.vdocuments.site/reader034/viewer/2022051912/600357b9e26afc0b2d54480b/html5/thumbnails/8.jpg)
Natural Bandwidth Boundaries
Bandwidthfrom a
processor’spoint of view
Span of memory
Chip (~10mm)
Board (~100mm)
Cabinet (~1000mm)
Floor
KBs TBs
![Page 9: MTAAP’07 Keynote - PNNLhpc.pnl.gov/mtaap/mtaap07/mtaap_files/keynote.pdf · 2014. 2. 7. · MTAAP’07 Keynote Michael Merrill. Outline What’s important... applications ... Processor](https://reader034.vdocuments.site/reader034/viewer/2022051912/600357b9e26afc0b2d54480b/html5/thumbnails/9.jpg)
Natural Bandwidth Boundaries
Bandwidthfrom a
processor’spoint of view
Span of memory
Chip (~10mm)
Board (~100mm)
Cabinet (~1000mm)
Floor
Near
Far
KBs TBs
Optics might merge Cabinet and
Floor levels
![Page 10: MTAAP’07 Keynote - PNNLhpc.pnl.gov/mtaap/mtaap07/mtaap_files/keynote.pdf · 2014. 2. 7. · MTAAP’07 Keynote Michael Merrill. Outline What’s important... applications ... Processor](https://reader034.vdocuments.site/reader034/viewer/2022051912/600357b9e26afc0b2d54480b/html5/thumbnails/10.jpg)
Trends we live with...
Performanceon a log scale
Time on a linear scale
Local Memory
ALU
Global MemoryDifficulty of Use
![Page 11: MTAAP’07 Keynote - PNNLhpc.pnl.gov/mtaap/mtaap07/mtaap_files/keynote.pdf · 2014. 2. 7. · MTAAP’07 Keynote Michael Merrill. Outline What’s important... applications ... Processor](https://reader034.vdocuments.site/reader034/viewer/2022051912/600357b9e26afc0b2d54480b/html5/thumbnails/11.jpg)
Different Balance
Costs have changed drastically
Transistors are cheap... Wires are expensive
Processor complexity vs. power is an issue
Balance costs... apply transistors to use wires more effectively... not just for cache
This is why you see architecture changing
![Page 12: MTAAP’07 Keynote - PNNLhpc.pnl.gov/mtaap/mtaap07/mtaap_files/keynote.pdf · 2014. 2. 7. · MTAAP’07 Keynote Michael Merrill. Outline What’s important... applications ... Processor](https://reader034.vdocuments.site/reader034/viewer/2022051912/600357b9e26afc0b2d54480b/html5/thumbnails/12.jpg)
What’s necessary... I think
Need to provide an effective system solution HW and SW!
Why? Days of coarse grained scaling are at an end... so threads/contexts will necessarily work together to perform a task.
![Page 13: MTAAP’07 Keynote - PNNLhpc.pnl.gov/mtaap/mtaap07/mtaap_files/keynote.pdf · 2014. 2. 7. · MTAAP’07 Keynote Michael Merrill. Outline What’s important... applications ... Processor](https://reader034.vdocuments.site/reader034/viewer/2022051912/600357b9e26afc0b2d54480b/html5/thumbnails/13.jpg)
Fabrication
Tiled architecture with partial good chips for lower costs
Detect failed computation
Retry failed computation
Move away from fixed number of threads/contexts
![Page 14: MTAAP’07 Keynote - PNNLhpc.pnl.gov/mtaap/mtaap07/mtaap_files/keynote.pdf · 2014. 2. 7. · MTAAP’07 Keynote Michael Merrill. Outline What’s important... applications ... Processor](https://reader034.vdocuments.site/reader034/viewer/2022051912/600357b9e26afc0b2d54480b/html5/thumbnails/14.jpg)
What’s necessary... I think
Multi-Context Processor
Compiler and Runtime
Coordination Synchronization
Need at least these three things working together to produce an effective environment for the application developer
![Page 15: MTAAP’07 Keynote - PNNLhpc.pnl.gov/mtaap/mtaap07/mtaap_files/keynote.pdf · 2014. 2. 7. · MTAAP’07 Keynote Michael Merrill. Outline What’s important... applications ... Processor](https://reader034.vdocuments.site/reader034/viewer/2022051912/600357b9e26afc0b2d54480b/html5/thumbnails/15.jpg)
Effective Software
Multi-Context Processor
Coordination Synchronization
Compiler and Runtime
Runtime that provides effective dynamic work management so the unbalanced nature of the application can be mitigated.
Compiler that takes advantage of such a runtime increases programmer effectiveness and productivity allowing them to concentrate on the application.
![Page 16: MTAAP’07 Keynote - PNNLhpc.pnl.gov/mtaap/mtaap07/mtaap_files/keynote.pdf · 2014. 2. 7. · MTAAP’07 Keynote Michael Merrill. Outline What’s important... applications ... Processor](https://reader034.vdocuments.site/reader034/viewer/2022051912/600357b9e26afc0b2d54480b/html5/thumbnails/16.jpg)
Latency Tolerance/Management
Multi-Context Processor
Coordination Synchronization
Compiler and Runtime
Effective use of the bandwidth provided by the internal system networks through the use of latency tolerance and/or latency management techniques.
Many of these techniques require the exposure of abundant fine-grained parallelism in the application.
![Page 17: MTAAP’07 Keynote - PNNLhpc.pnl.gov/mtaap/mtaap07/mtaap_files/keynote.pdf · 2014. 2. 7. · MTAAP’07 Keynote Michael Merrill. Outline What’s important... applications ... Processor](https://reader034.vdocuments.site/reader034/viewer/2022051912/600357b9e26afc0b2d54480b/html5/thumbnails/17.jpg)
Low Overhead Coordination
Multi-Context Processor
CoordinationSynchronization
Compiler and Runtime
Threads will necessarily work together to compute so effective coordination will be essential.
Any cycles spent waiting on synchronization events are not spent computing and therefore decrease efficiency.
![Page 18: MTAAP’07 Keynote - PNNLhpc.pnl.gov/mtaap/mtaap07/mtaap_files/keynote.pdf · 2014. 2. 7. · MTAAP’07 Keynote Michael Merrill. Outline What’s important... applications ... Processor](https://reader034.vdocuments.site/reader034/viewer/2022051912/600357b9e26afc0b2d54480b/html5/thumbnails/18.jpg)
What’s possible... maybe
Don’t look for any major companies to make things significantly better because it messes with the current business too much.Which direction to go?
![Page 19: MTAAP’07 Keynote - PNNLhpc.pnl.gov/mtaap/mtaap07/mtaap_files/keynote.pdf · 2014. 2. 7. · MTAAP’07 Keynote Michael Merrill. Outline What’s important... applications ... Processor](https://reader034.vdocuments.site/reader034/viewer/2022051912/600357b9e26afc0b2d54480b/html5/thumbnails/19.jpg)
Straight Forward Scaling
90nm 65nm 45nm 32nmTU 160 306 640 1266
FPU 80 153 320 633TU/XB 2 3 4 6XBar 80 102 160 211Clock 500M 585M 684M 800MPerf 80G 179G 437G 1.01T
SRAM 4.8M 9.2M 19.2M 37.9M
start with Cyclops64
22x23mm die
150W to 190W
3DE ?
![Page 20: MTAAP’07 Keynote - PNNLhpc.pnl.gov/mtaap/mtaap07/mtaap_files/keynote.pdf · 2014. 2. 7. · MTAAP’07 Keynote Michael Merrill. Outline What’s important... applications ... Processor](https://reader034.vdocuments.site/reader034/viewer/2022051912/600357b9e26afc0b2d54480b/html5/thumbnails/20.jpg)
What About Software?
Need good compiler technology to exploit on chip explicit memoryMuch higher level of abstractionNeed to separate the how-to from the what-for but express bothDiagnose hot spots (resource contention)etc...
![Page 21: MTAAP’07 Keynote - PNNLhpc.pnl.gov/mtaap/mtaap07/mtaap_files/keynote.pdf · 2014. 2. 7. · MTAAP’07 Keynote Michael Merrill. Outline What’s important... applications ... Processor](https://reader034.vdocuments.site/reader034/viewer/2022051912/600357b9e26afc0b2d54480b/html5/thumbnails/21.jpg)
Questions? ... I have a ton ;-)