Computer Systems are Different! Frans Kaashoek and Robert Morris 6.033 Spring 2009Composibility via static discipline • Be tolerant of inputs and strict on outputsMoore’s law “Cramming More Components Onto Integrated Circuits”, Electronics, April 1965 cost per transistor transistors per dieTransistors/die doubles every ~18 monthsLithography: the driver behind transistor count • Components/area O(x2) with feature size • Total components O(a) with die area • Switching rate O(x) with feature sizeCPU performanceDRAM densityDisk: Price per GByte drops at ~30-35% per yearENIAC • 1946 • Only one • 5000 adds/sec • 20 10-digit registers • 18,000 vacuum tubes • 124,500 watts • Not really stored programUNIVAC (Universal Automatic Computer) • 1951 • 46 sold • 2000 ops/sec • 1,000 12-digit words (mercury) • 5000 tubes • $1.5 millionIBM System/360-40 • 1964 • 1.6 MHz • 16-256 KB core • $225,000 • Family of six • 32-bit • Time-sharingCray 1: supercomputer • 1976 • 80 sold • 80 MHz • 8 Mbyte SRAM • 230,000 gates • $5 millionDEC PDP-8 (1964) • 60,000 sold • 330,000 adds/sec • 4096 12-bit words • $18,000Apple II • 1977 • 1 MHz • 6502 microprocessor • 4 to 48 Kilobytes RAM • $1300 • Basic, VisicalcIBM’s wrist watch • 2001 • Linux and X11 • 74 Mhz CPU • 8 Megabyte flash • 8 Megabyte DRAM • WirelessSoftware follows hardware 0102030405060Windows 3.1 (1992)Windows NT (1992)Solaris (1998)Windows 95Windows 98Windows NT 5.0 (1998)RedHat Linux 6.2 (2000)RedHat Linux 7.1 (2001)Windows XPVistaMillions of lines of source codeCheap → PervasivePervasive → qualitative change year log (people per computer) Slide from David Culler, UC Berkeley Number crunching Embedded Sense/control Word processing CommunicationLatency improves slowly 11010010001 2 3 4 5 6 7 8 9 10 11Year # Improvement wrt year #1 Moore’s law (~70% per year) DRAM access latency (~7% per year) Speed of light (0% per year)Heat is a problemRecent Intel CPU Clock Rates 486 Pentium PentiumPro Pentium III Pentium 4 Pentium 4 HT mHzThe Future: will it be painful? AMD Barcelona Quad-core chipWhat went right? • Unbounded composibility • General-purpose computers • Only need to make one thing fast • Separate arch from implementation • S/W can exploit new H/W • Cumulative R&D investment over
View Full Document