+ CS263: Runtime Systems

Transcription

1 + CS263: Runtime Systems Spring Prof. Chandra Krintz Laboratory for Research on Adaptive Computing Environments (RACELab) Computer Science Department Harold Frank Hall (HFH) 2153

2 + Past 40 Years in Technology Were Extraordinary 2 n Sustained exponential improvement in fundamental technologies n Exponential: gain in 2 years = all gains overall previous years n 1974 n computers used by business n 2014 n computer in every home n information everywhere, anytime n computer in every pocket n computer in every object

3 + Unprecedented Technology Change 3 Unprecedented Technology Change

4 + Moore s Law 4

5 + Moore s Law In Practice Intel transistors 740 khz clock 10um process 10.8 usec/inst Intel Core i7 980X 1.17B transistors 3.33 GHz clock 32nm process 73.4 psec/inst Improvement/year Ratio 38% % % %

6 + Moore s Secret: Dennard Scaling 6 Design of Ion-Implanted MOSFET S with Very Small Physical Dimensions ROBERT H. DENNARD, LIEMBER, IEEE, FRITZ H. GAENSSLEN, HWA-NIEN YU, MEMBER, IEEE, V. LEO RIDEOUT, MEMBER) IEEE, ERNEST BASSOUS, AND ANDRE R. LEBLANC, MEMBER, IEEE paper considers the design, fabrication, and LIST OF SYMBOLS Absfracf This characterization of very small MOSI?ET switching devices suitable Inverse semilogarithmic slope of subthreshold a for digital integrated circuits using dimensions of the order of 1 p. characteristic. Scaling relationships are presented which show how a conventional D Width of idealized step function profde MOSFET can be reduced in size. An improved small device structure is presented that uses ion implantation to provide shallow for chadnel implant. source and drain regions and a nonuniform substrate doping profile. AW, Work function difference between gate One-dimensional models are used to predict the substrate and substrate. doping profile and the corresponding threshold voltage versus source voltage characteristic. A two-dimensional current transport Dielectric constants for silicon and model is used to predict the relative degree of short-channel effects silicon dioxide. for different device parameter combinations. Polysilicon-gate Drain current. MOSFET S with channel lengths as short as 0.5 ~ were fabricated, Boltzmann s constant. and the device characteristics measured and compared with predicted Unitless scaling constant. values. The performance improvement expected from using these very small devices in highly miniaturized integrated circuits MOSFET channel length. is projected. Effective surface mobility. Intrinsic carrier concentration. Substrate acceptor concentration. Band bending in silicon at the onset of Manuscript received May 20, 1974; revised July 3, strong inversion for zero substrate The aubhors are with the IBM T. J. Watson Research Center, Yorktown Heights, N.Y voltage. Device or Circuit Parameter Scaling Factor Dimension, Tox, L, W 1/k Doping Concentration Na k Voltage (V) 1/k Current (I) 1/k Capacitance (ea/t ) 1/k Delay time/circuit (VC/I) 1/k Power dissipation/circuit (VI) 1/k^2 Power density (VI/A) 1 Historically, k ~= 1.4 ( 2) [Dennard, Gaensslen, Yu, Rideout, Bassous, Leblanc, IEEE JSSC, 1974]

7 + Dennard Scaling is Dead 7

8 + That Was Fun! What s Next? 8

9 + Traditional Sources of Improvement 9 Compilers Computer Architecture Semiconductors

10 + New Opportunities 10 Reconfigurable Computing Distributed Systems Software

11 + State of Software 11 n Software is large, complex, and bloated n Emphasis on programmer productivity, not software efficiency n Performance improvement opportunities abound n Not long-term, secular trend like Moore s Law, but still important

12 + Large & Bloated Ex: Linux Growth Size (Linux 1.0 = 1) Linux Size Moores Law 1 1/31/ /28/1995 7/24/1998 4/19/2001 1/14/ /10/2006 7/6/2009 4/1/ /27/2014

13 + Large & Bloated Ex: Linux Complexity 13

14 + Large & Bloated Ex: Windows Growth 14 Recommended Minimum Configuration (32 bit) Windows 3.1 Windows 95 Windows 98 Windows 2000 Windows XP Vista Premium Windows 7 Processor (SPECInt) Memory (MB) Disk (MB) Moore's Law

15 + Software Bloat 15 Transform date from SOAP message to Java object (IBM Trade benchmark) 268 calls 70 objects allocated

16 + Computer Science is the Science of Abstraction 16

17 + Object Bloat 17 Array holding 1 string Hash set containing 3 strings Primitive Header Pointer Null % 55.6% 11.1% 11.1% Primitive Header Pointer Null % 42.9% 11.0% 21.2%

18 + What is Going On? 18 Developer Efficiency Language Inefficiency Frameworks Systems Abstraction

19 + Developer Efficiency 19 n Time to market valued over execution efficiency n First mover advantage in competitive world n Features more important than memory footprint or execution time n High-level languages and rich libraries n Modularity and abstraction essential to develop complex codes

20 + 20 -Unmanaged language: statically compiled, architecture-dependent binary, streamlined runtime (C, C++,VB,asm,ObjC/Swift,Go) -Managed: high-level, architecture-independent (portable) binary format, runtime performs translation (all others) From: Tiobe 2017 Roughly equivalent to number of lines of code in the wild

21 + 21 Coding Dojo 2017 From job postings

22 + Hello World! ,000 Avg. Ticks (280ns) Lower is better 10,000 1, C Console C Window C# Console C# Window 1 Hello World

23 + Language Implementations: Runtime Systems/VMs 23 n Collection of general-purpose components, libraries, frameworks n Java,.NET, WebSphere, n Productivity through reuse of high quality, high-level abstractions n Flipside of generality is inefficiency n Appeal to widest audience by handling many scenarios n Bloated, complex software n Unused functionality tax n Not specialized to specific use n Cut/pasted solutions restrict true understanding and introduce bugs

24 + Abstraction is Bad (For Performance) 24 n Abstraction captures functionality, obscures performance n Performance characteristic of implementation, not interface n Performance tuning destroys abstraction boundaries n But, abstraction essential to construct large, complex systems n Cannot understand or predict performance of these systems n Little work on specifying, analyzing, or modeling performance n Big-O notation hides too much n Compilers and parallelism have not been able to solve the problem

25 + Are Languages or Runtimes the Problem? 25 n Type safe, memory safe, modern programming languages n Not necessarily intrinsically expensive: MSR Singularity OS in C# n But, some very popular languages have very poor implementations n Portability over performance (interpreter only) n Global interpreter lock in Python ( n Dynamic typing = Run-time checks + barrier to compiler optimization n Unsophisticated compilers in widely used implementations, if even available n Lack support for emerging software needs n High-performance, distributed computing/asynchrony, concurrency/parallelism, scalability, data-oriented computing, reliability,...

26 + Matrix Multiply Saman Amarasinghe, MIT Fall 2011

27 + Increasing Performance of Software and Systems 27 n Managed runtime systems have become the norm n Requires that we understand what they do, what they hide, how they work, and how they can be improved n Performance implications of n Object orientation n Garbage collection n How managed runtime system (VMs for high-level languages) work n Interpretation n Compilation (dynamic and JIT) n Performance monitoring n Adaptive optimization

28 + CS n Managed runtime systems have become the norm n Requires that we understand what they do, what they hide, how they work, and how they can be improved n Performance implications of n Object orientation n Garbage collection n How managed runtime system (VMs for high-level languages) work n Interpretation n Compilation (dynamic and JIT) n Performance monitoring n Adaptive optimization

29 + CS n Managed runtime systems have become the norm n Requires that we understand what they do, what they hide, how they work, and how they can be improved n Performance implications of n Object orientation n Garbage collection n How managed runtime system (VMs for high-level languages) work n Interpretation n Compilation (dynamic and JIT) n Performance monitoring n Adaptive optimization

30 + CS263 Evaluation 30 n 50% Homework assignments and quizzes in and out of class n Includes class participation; no makeups or date changes given n 50% Project (2 person groups) n Weekly code commits (starting week 2) n Public github repo, documentation to build/regenerate n minute in class presentation and demo the last week(s) of class n 5 page writeup n Problem, solution, evaluation n Project ideas posted on web page

31 + Questions? 31 n Instructor: Chandra Krintz n HFH 2153 n Office hours by appointment, skype (ckrintz), chat (ckrintz@gmail.com, ckrintz) n Webpage n Lectures posted (slides and youtube) n Class starts promptly at 9am (please be on time) n Will end between 10:15 and 10:45 depending on the topic n Assigned readings on website/schedule should be read by the class date indicated

32 32