Computer Systems Structure Main Memory Organization Peripherals Computer Central Main Processing Memory Unit Computer Systems Interconnection Input Output Communication lines CS 160 Ward 1 CS 160 Ward 2 Storage/Memory Hierarchy Storage Hierarchy & Characteristics CS 160 Ward 3 CS 160 Ward 4 Principle of Memory Hierarchy Memory Hierarchy Importance 1980: no cache in µproc; 1995: 2-level cache on chip To optimize memory performance for a given (1989 first Intel µproc with a cache on chip) cost, a set of technologies are arranged in a 1000 CPU hierarchy that contains a relatively small amount of fast memory and larger amounts of less expensive, but slower memory. 100 Processor-Memory Performance Gap: (grows 50% / year) 10 Performance DRAM 1 1990 1995 1980 1985 2000 Time CS 160 Ward 5 CS 160 Ward 6 Storage Characteristics Location • Location • CPU • Capacity • Internal • Unit of transfer • External • Access method • Performance • Physical type • Physical characteristics • Organization CS 160 Ward 7 CS 160 Ward 8 Capacity Unit of Transfer • Word size • Internal – The natural unit of organization – Usually governed by data bus width – Expected size of most data & instructions • External – Typically 32 bits or 64 bits – Usually a block which is much larger than a word • Past: 16 bits • Addressable unit • Number of words – Smallest unit which can be uniquely addressed – or Bytes – Byte internally (typically) CS 160 Ward 9 CS 160 Ward 10 Access Methods (1) Access Methods (2) • Sequential (e.g., tape) – Shared read/write mechanism • Random (e.g., RAM) – Start at the beginning and read through in order – Individual addresses identify locations exactly – Access time depends on location of data and previous – Access time is independent of location or previous location access • Direct (e.g., disk) • Associative (e.g., cache) – Shared read/write mechanism – Data is located by a comparison with contents of a – Individual blocks have unique address portion of the store – Access is by jumping to vicinity plus sequential – Access time is independent of location or previous search access – Access time depends on location and previous location CS 160 Ward 11 CS 160 Ward 12 Performance Transfer Rate Example Problem • Latency/Access time • Assume we have 32-Mbit DRAM memory with 8 – Time between presenting the address and getting the bits simultaneously read and a cycle time 250 ns. valid data (e.g., in memory, time between the Read & Memory Function Competed (MFE) signals) • How fast can data be moved out of memory (e.g. • Memory Cycle time transfer rate)? – Time may be required for the memory to “recover” before next access 8b*(1/250ns) – Cycle time is latency + recovery = 8b*(4x106/s) • Transfer Rate – Rate at which data can be moved = 32 Mbps – (# of bits(bytes)) * (1/(cycle time)) = 4 MBps CS 160 Ward 13 CS 160 Ward 14 Physical Types Physical Characteristics • Semiconductor – RAM • Volatility • Magnetic • Erasable – Disk & Tape • Power consumption / Heat • Optical – CD & DVD • Others – Bubble – Hologram CS 160 Ward 15 CS 160 Ward 16 Organization The Bottom Line • How much? • Physical arrangement of bits into words – Capacity • Not always obvious • How fast? • e.g. interleaved (striped) – Performance (Time is money) • How expensive? CS 160 Ward 17 CS 160 Ward 18 Hierarchy List • Registers • L1 Cache (on-chip) • L2 Cache (off-chip) • Memory controller cache (L3 Cache) Memory Basics • Main memory • Disk cache • Disk • Optical • Tape CS 160 Ward 19 CS 160 Ward 20 Main Memory Basics Data Flow (Indirect Diagram) • Memory: where computer stores programs and data • Bit (binary digit): basic unit. (8 bits = byte) • Each memory cell (location) has an address numbered 0, …, n -1 (for n memory cells) • Possible address range limited by address size (m bits in address means 2m addresses) • Memory cell size (typically 1 byte) grouped together into words (typically 32 or 64 bits) • 32-bit machine will typically have 32-bit registers and instructions for manipulating 32-bit words; similarly for 64-bit machine CS 160 Ward 21 CS 160 Ward 22 Semiconductor Memory Static RAM • Random Access Memory (RAM) • Bits stored as on/off switches (transistors) – All semiconductor memory is random access • No charges to leak (directly accessed via address logic) • No refreshing needed when powered – Read/Write • Larger per bit – Volatile (requires constant power supply) • More expensive – Temporary storage • Does not need refresh circuits – Static (holds data) • Faster or dynamic (periodically refreshes charge) • Example: – Cache CS 160 Ward 23 CS 160 Ward 24 SRAM Illustration Dynamic RAM • Bits stored as charge in capacitors (also uses transistors) – Charges leak – Need refreshing even when powered • Smaller per bit • Less expensive • Need refresh circuits • When enable is high, output is same as input. • Slower • Otherwise, output holds last value • Asynchronous and Synchronous DRAMs • Example: – Main memory CS 160 Ward 25 CS 160 Ward 26 DRAM Illustration Read Only Memory (ROM) • Permanent storage • Microprogramming • Library subroutines • Systems programs • Function tables • More complex than figure implies • Must coordinate with normal read and write operations CS 160 Ward 27 CS 160 Ward 28 Measures of Memory Technology Memory Density • Refers to memory cells per square area of silicon • Density • Usually stated as number of bits on standard • Latency and cycle time chip size • Examples: – 1 meg chip holds 1 megabit of memory – 4 meg chip holds 4 megabit of memory • Memory cells typically in arrays – 1M x 1 chip is a 1 meg chip – 256K x 4 chip is a 1 meg chip • Note: higher density chip generates more heat CS 160 Ward 29 CS 160 Ward 30 Internal Module Organization [1] Internal Module b7 b′7 b1 b′1 b0 b′0 W0 • • • Organization [2] FF FF A 0 W 1 • • • A 1 Address • • • • • • Memory decoder cells A • • • • • • 2 • • • • • • A 3 W15 • • • Compare with Fig 5.3 in textbook Sense / Write Sense / Write Sense / Write R/W circuit circuit circuit CS 256K x 8 chip Data input/output lines: b7 b1 b0 16 x 8 chip CS 160 Ward 31 CS 160 Ward 32 Internal Module Organization [3] Typical 16 Mb DRAM (Internal) 1M x 8 chip 4M x 4 chip CS 160 Ward 33 CS 160 Ward 34 Memory Packaging: Chips Read-Write Performance • 16-Mbit chip (4M x 4) • In many memory technologies, the time 4M in 11 rows by 11 columns (222=4M) required to fetch information from memory WE = write enable differs from the time required to store OE = output enable information in memory, and the difference RAS = row address select can be dramatic. Therefore, any measure of CAS = column address select memory performance must give two values: the performance of operations and the A0 – A10 = 11 address bits read performance of write operations. D1 – D4 = Data to be read/written NC = no connect (even # of pins) Vcc = power supply Vss = ground pin CS 160 Ward 35 CS 160 Ward 36 Memory Organization Memory Transfer • Memory controller connects computer Physical memory is organized into words, where to physical memory chips a word is equal to the memory transfer size. Each read and write operation applies to an entire word. Remember: The physical memory hardware does not • Latency provide a way to read or write less than a complete word. • Cycle time (read and write) Example: • Transfer size (or 32 bits = word word size) CS 160 Ward 37 CS 160 Ward 38 Byte & Word Addresses Address Translation • CPU can use byte addresses (convenient) To avoid arithmetic calculations, such as division or • Physical memory can use word addresses remainder, physical memory is organized such (efficient) that the number of bytes per word is a power of • Translation performed by intelligent memory two, which means the translation from byte controller address to a word address and offset can be performed by extracting bits. CS 160 Ward 39 CS 160 Ward 40 Performance Enhancements Illustration of Interleaving • Memory Banks – Alternative to single memory and single memory controller – Processor connects to multiple controllers – Each controller connects to separate physical memory – Controllers and memory can operate simultaneously • Interleaving – Related to memory banks – Transparent to programmer – Places consecutive bytes in separate physical memories – Uses low-order bits of address to choose module Consecutive bytes stored in separate physical memories. – Known as N-way interleaving (N = number of physical memories) CS 160 Ward 41 CS 160 Ward 42 Error Correction for Data • Semiconductor Memory – Hard Failure • Permanent defect • Caused by Error Correction – Environmental abuse – Manufacturing defects – Wear – Soft Error • Random, non-destructive • No permanent damage to memory • Caused by – Voltage spikes – Alpha particles • Data Transmission CS 160 Ward 43 CS 160 Ward 44 Parity as Error Detector Error-Correcting Codes • Parity can check single bit errors – Store one extra bit (parity bit) for each word – Even parity: have parity bit set so even number of 1’s • 10010101: ok • 10000101: not ok (some bit is wrong, don’t know which one) – Odd parity: have parity bit set so odd number of 1’s CS 160 Ward 45 CS 160 Ward 46 4-Bit Hamming Error-Correcting Code Check Bits for Error Correction CS 160 Ward 47 CS 160 Ward 48 Single-Bit Error Detection/Correction 8-Bit Error: Check Bit Calculation • Hamming code for any size memory word – Given r check/parity bits and m data bits, word size = m + r bits – Number bits from right to left starting at 1 ( not 0) – All bits with power of 2 number are check bits (bits C1 = D1 ⊕ D2 ⊕ D4 ⊕ D5 ⊕ D7 numbered 1, 2, 4, 8, 16, … )