Presentation is loading. Please wait.

Presentation is loading. Please wait.

21st Century Computer Architecture A community white paper 21st Century Computer Architecture.

Similar presentations


Presentation on theme: "21st Century Computer Architecture A community white paper 21st Century Computer Architecture."— Presentation transcript:

1 21st Century Computer Architecture A community white paper 21st Century Computer Architecture A community white paper Technion, Haifa Israel, June 2013 Information & Commun. Tech’s Impact Semiconductor Technology’s Challenges Computer Architecture’s Future Example: Bypassing Paged Virtual Memory

2 White Paper Participants “*” contributed prose; “**” effort coordinator Thanks of CCC, Erwin Gianchandani & Ed Lazowska for guidance and Jim Larus & Jeannette Wing for feedback 2 Sarita Adve, U Illinois * David H. Albonesi, Cornell U David Brooks, Harvard U Luis Ceze, U Washington * Sandhya Dwarkadas, U Rochester Joel Emer, Intel/MIT Babak Falsafi, EPFL Antonio Gonzalez, Intel/UPC Mark D. Hill, U Wisconsin *,** Mary Jane Irwin, Penn State U * David Kaeli, Northeastern U * Stephen W. Keckler, NVIDIA/U Texas Christos Kozyrakis, Stanford U Alvin Lebeck, Duke U Milo Martin, U Pennsylvania José F. Martínez, Cornell U Margaret Martonosi, Princeton U * Kunle Olukotun, Stanford U Mark Oskin, U Washington Li-Shiuan Peh, M.I.T. Milos Prvulovic, Georgia Tech Steven K. Reinhardt, AMD Michael Schulte, AMD/U Wisconsin Simha Sethumadhavan, Columbia U Guri Sohi, U Wisconsin Daniel Sorin, Duke U Josep Torrellas, U Illinois * Thomas F. Wenisch, U Michigan * David Wood, U Wisconsin * Katherine Yelick, UC Berkeley/LBNL *

3 20 th Century ICT Set Up Information & Communication Technology (ICT) Has Changed Our World o Required innovations in algorithms, applications, programming languages, …, & system software Key (invisible) enablers (cost-)performance gains o Semiconductor technology (“Moore’s Law”) o Computer architecture (~80x per Danowitz et al.) 3

4 Enablers: Technology + Architecture 4 Danowitz et al., CACM 04/2012, Figure 1 Technology Architecture

5 21 st Century Promise ICT Promises Much More o Data-centric personalized health care o Computation-driven scientific discovery o Human network analysis o Much more: known & unknown Characterized by o Big Data o Always Online o Secure/Private o…o… Whither enablers of future (cost-)performance gains? 5

6 Technology’s Challenges 1/2 Late 20 th CenturyThe New Reality Moore’s Law — 2× transistors/chip Transistor count still 2× BUT… Dennard Scaling — ~constant power/chip Gone. Can’t repeatedly double power/chip 6

7 Classic CMOS Dennard Scaling: the Science behind Moore’s Law 7 National Research Council (NRC) – Computer Science and Telecommunications Board (CSTB.org) Scaling: Oxide: t OX /  Results: Power Density: Voltage: V/  Power/ckt: 1/  2 ~Constant (Finding 2) Source: Future of Computing Performance: Game Over or Next Level?, National Academy Press, 2011

8 Power Density: ~Constant Post-classic CMOS Dennard Scaling 8 National Research Council (NRC) – Computer Science and Telecommunications Board (CSTB.org) Scaling: Oxide: t OX /  Results: Voltage: V/  V Power/ckt: 1 22 1/  2 Post Dennard CMOS Scaling Rule TODO: Chips w/ higher power (no), smaller (  ), dark silicon ( ), or other (?)

9 Technology’s Challenges 2/2 Late 20 th CenturyThe New Reality Moore’s Law — 2× transistors/chip Transistor count still 2× BUT… Dennard Scaling — ~constant power/chip Gone. Can’t repeatedly double power/chip Modest (hidden) transistor unreliability Increasing transistor unreliability can’t be hidden Focus on computation over communication Communication (energy) more expensive than computation 1-time costs amortized via mass market One-time cost much worse & want specialized platforms 9 How should architects step up as technology falters?

10 21 st Century Comp Architecture 20 th Century21 st Century Single-chip in generic computer Architecture as Infrastructure: Spanning sensors to clouds Performance plus security, privacy, availability, programmability, … Cross- Cutting: Break current layers with new interfaces Performance via invisible instr.-level parallelism Energy First ● Parallelism ● Specialization ● Cross-layer design Predictable technologies: CMOS, DRAM, & disks New technologies (non-volatile memory, near-threshold, 3D, photonics, …) Rethink: memory & storage, reliability, communication 10 X X

11 21 st Century Comp Architecture 20 th Century21 st Century Single-chip in stand-alone computer Architecture as Infrastructure: Spanning sensors to clouds Performance plus security, privacy, availability, programmability, … Cross- Cutting: Break current layers with new interfaces Performance via invisible instr.-level parallelism Energy First ● Parallelism ● Specialization ● Cross-layer design Predictable technologies: CMOS, DRAM, & disks New technologies (non-volatile memory, near-threshold, 3D, photonics, …) Rethink: memory & storage, reliability, communication 11

12 What Research Exactly? Research areas in white paper (& backup slides) 1.Architecture as Infrastructure: Spanning Sensors to Clouds 2.Energy First 3.Technology Impacts on Architecture 4.Cross-Cutting Issues & Interfaces Much more research developed by future PIs! E.g.: Efficient Virtual Memory for Big Memory Servers o Basu, Gandhi, Chang, Hill, & Swift [ISCA 2013] o Big Memory: graph500, memcached, databases o Self-manage most memory (e.g., bufferpool) 12

13 10/5/1213 Execution Time Overhead: TLB Misses 1.Significant waste 2.Larger memory? 3.Byte-addr NVM? Lower is better

14 Hardware: Direct Segment OFFSET BASE LIMIT VA Conventional Paging PA Direct Segment Why Direct Segment? Matches Big Memory Workload needs NO Paging => NO TLB Miss

15 Execution Time Overhead: TLB Misses 10/5/ % TLB “misses” to direct segment Requires: Both small SW + small HW changes

16 21st Century Computer Architecture A community white paper 21st Century Computer Architecture A community white paper Technion, Haifa Israel, June 2013 Information & Commun. Tech’s Impact Semiconductor Technology’s Challenges Computer Architecture’s Future Example: Bypassing Paged Virtual Memory

17 Pre-Competitive Research Justified Retain (cost-)performance enabler to ICT revolution Successful companies cannot do this by themselves o Lack needed long-term focus o Don’t want to pay for what benefits all o Resist transcending interfaces that define their products 17

18 White Paper Process Late March 2012 o CCC contacts coordinator & forms group April 2012 o Brainstorm (meetings/online doc) o Read related docs (PCAST, NRC Game Over, ACAR1/2, …) o Use online doc for intro & outline then parallel sections o Rotated authors to revise sections May 2012 o Brainstorm list of researcher in/out of comp. architecture o Solicit researcher feedback/endorsement o Do distributed revision & redo of intro o Release May 25 to CCC & via Kudos to participants on executing on a tight timetable 18

19 Back Up Slides Detailed research areas in white paper 1.Architecture as Infrastructure: Spanning Sensors to Clouds 2.Energy First 3.Technology Impacts on Architecture 4.Cross-Cutting Issues & Interfaces Findings on National Academy “Game Over” Study Glimpse at DARPA/ISAT Workshop “Advancing Computer Systems without Technology Progress” 19

20 1. Architecture as Infrastructure: Spanning Sensors to Clouds Beyond a chip in a generic computer To pillar of 21 st century societal infrastructure. o Computation in context (sensor, mobile, …, data center) o Systems often large & distributed o Communication issues can dominate computation o Goals beyond performance (battery life, form factor) Opportunities (not exhaustive) o Reliable sensors harvesting (intermittent) energy o Smart phones to Star Trek’s medical “tricorder” o Cloud infrastructure suitable for both “Big Data” streams & low-latency qualify-of-service with stragglers o Analysis & design tools that scale 20

21 2. Energy First Beyond single-core performance computer To (cost-)performance per watt/joule Energy across the layers o Circuit/technology (near-threshold CMOS, 3D stacking) o Architecture (reducing unnecessary data movement) o Software (communication-reducing algorithms) Parallelism to save energy o Vast (fined-grained) homogeneous & heterogeneous o Improved SW stack o Applications focus (beyond graphic processing units) Specialization for performance & energy efficiency o Abstractions for specialization (reducing 1-time cost) o Energy-efficient memory hierarchies o Reconfigurable logic structures 21

22 3. Technology Impacts on Architecture Beyond CMOS, Dram, & Disks of last 3+ decades to Using replacement circuit technologies o Sub/near-threshold CMOS, QWFETs, TFETs, and QCAs Non-volatile storage o Beyond flash memory to STT-RAM, PCRAM, & memristor 3D die stacking & interposers o logic, cache, small main memory Photonic interconnects o Inter- & even intra-chip Design automation o from circuit-design w/ new technologies to o pre-RTL functional, performance, power, area modeling of heterogeneous chips & systems 22

23 4. Cross-Cutting Issues & Interfaces Beyond performance w/ stable interfaces to New design goals (for pillar of societal infrastructure) o Verifiability (bugs kill) o Reliability (“dependability” computing base?) o Security/Privacy (w/ non-volatile memory?) o Programmability (time to correct-performant solution) Better Interfaces o High-level information (quality of service, provenance) o Parallelism ((in)dependence, (lack of) side-effects) o Orchestrating communication ((recursive) locality) o Security/Reliability (fine-grain protection) 23

24 Executive summary (Added to National Academy Slides)  Highlights of National Academy Findings (F1) Computer hardware has transitioned to multicore (F2) Dennard scaling of CMOS has broken down (F3) Parallelism and locality must be exploited by software (F4) Chip power will soon limit multicore scaling  Eight recommendations from algorithms to education  We know all of this at some level, BUT: Are we all acting on this knowledge or hoping for business as usual? Thinking beyond next paper to where future value will be created? – Questions Asked but Not Answered Embedded in NA Talk – Briefly Close with Kübler-Ross Stages of Grief: Denial  …  Acceptance Source: Future of Computing Performance: Game Over or Next Level?, National Academy Press, 2011 Mark Hill talk ( )

25 The Graph 25 System Capability (log) 80s90s00s10s20s30s40s CMOS Fallow Period New Technology Our Focus 50s Source: Advancing Computer Systems without Technology Progress, ISAT Outbrief (http://www.cs.wisc.edu/~markhill/papers/isat2012_ACSWTP.pdf) Mark D. Hill and Christos Kozyrakis, DARPA/ISAT Workshop, March 26-27, Approved for Public Release, Distribution Unlimited The views expressed are those of the author and do not reflect the official policy or position of the Department of Defense or the U.S. Government.

26 Surprise 1 of 2 Can Harvest in the “Fallow” Period! 2 decades of Moore’s Law-like perf./energy gains Wring out inefficiencies used to harvest Moore’s Law HW/SW Specialization/Co-design (3-100x) Reduce SW Bloat (2-1000x) Approximate Computing (2-500x) ~1000x = 2 decades of Moore’s Law! 26

27 “Surprise” 2 of 2 Systems must exploit LOCALITY-AWARE parallelism Parallelism Necessary, but not Sufficient As communication’s energy costs dominate Shouldn’t be a surprise, but many are in denial Both surprises hard, requiring “vertical cut” thru SW/HW 27


Download ppt "21st Century Computer Architecture A community white paper 21st Century Computer Architecture."

Similar presentations


Ads by Google