Download Computer Systems: A Programmer's Perspective - Memory Hierarchies and Caching and more Study notes Geometry in PDF only on Docsity! Carnegie Mellon 1 Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspec;ve, Third Edi;on The Memory Hierarchy 15-‐213: Introduc;on to Computer Systems 11th Lecture, Oct. 6, 2015 Instructors: Randal E. Bryant and David R. O’Hallaron Carnegie Mellon 2 Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspec;ve, Third Edi;on Today ¢ Storage technologies and trends ¢ Locality of reference ¢ Caching in the memory hierarchy Carnegie Mellon 5 Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspec;ve, Third Edi;on NonvolaCle Memories ¢ DRAM and SRAM are volaCle memories § Lose informa;on if powered off. ¢ NonvolaCle memories retain value even if powered off § Read-‐only memory (ROM): programmed during produc;on § Programmable ROM (PROM): can be programmed once § Eraseable PROM (EPROM): can be bulk erased (UV, X-‐Ray) § Electrically eraseable PROM (EEPROM): electronic erase capability § Flash memory: EEPROMs. with par;al (block-‐level) erase capability § Wears out aZer about 100,000 erasings ¢ Uses for NonvolaCle Memories § Firmware programs stored in a ROM (BIOS, controllers for disks, network cards, graphics accelerators, security subsystems,…) § Solid state disks (replace rota;ng disks in thumb drives, smart phones, mp3 players, tablets, laptops,…) § Disk caches Carnegie Mellon 6 Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspec;ve, Third Edi;on TradiConal Bus Structure ConnecCng CPU and Memory ¢ A bus is a collecCon of parallel wires that carry address, data, and control signals. ¢ Buses are typically shared by mulCple devices. Main memory I/O bridge Bus interface ALU Register file CPU chip System bus Memory bus Carnegie Mellon 7 Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspec;ve, Third Edi;on Memory Read TransacCon (1) ¢ CPU places address A on the memory bus. ALU Register file Bus interface A 0 A x Main memory I/O bridge %rax Load operation: movq A, %rax Carnegie Mellon 10 Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspec;ve, Third Edi;on Memory Write TransacCon (1) ¢ CPU places address A on bus. Main memory reads it and waits for the corresponding data word to arrive. y ALU Register file Bus interface A Main memory 0 A %rax I/O bridge Store operation: movq %rax, A Carnegie Mellon 11 Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspec;ve, Third Edi;on Memory Write TransacCon (2) ¢ CPU places data word y on the bus. y ALU Register file Bus interface y Main memory 0 A %rax I/O bridge Store operation: movq %rax, A Carnegie Mellon 12 Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspec;ve, Third Edi;on Memory Write TransacCon (3) ¢ Main memory reads data word y from the bus and stores it at address A. y ALU Register file Bus interface y main memory 0 A %rax I/O bridge Store operation: movq %rax, A Carnegie Mellon 15 Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspec;ve, Third Edi;on Disk Geometry (Muliple-‐PlaXer View) ¢ Aligned tracks form a cylinder. Surface 0 Surface 1 Surface 2 Surface 3 Surface 4 Surface 5 Cylinder k Spindle Platter 0 Platter 1 Platter 2 Carnegie Mellon 16 Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspec;ve, Third Edi;on Disk Capacity ¢ Capacity: maximum number of bits that can be stored. § Vendors express capacity in units of gigabytes (GB), where 1 GB = 109 Bytes. ¢ Capacity is determined by these technology factors: § Recording density (bits/in): number of bits that can be squeezed into a 1 inch segment of a track. § Track density (tracks/in): number of tracks that can be squeezed into a 1 inch radial segment. § Areal density (bits/in2): product of recording and track density. Carnegie Mellon 17 Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspec;ve, Third Edi;on Recording zones ¢ Modern disks parCCon tracks into disjoint subsets called recording zones § Each track in a zone has the same number of sectors, determined by the circumference of innermost track. § Each zone has a different number of sectors/track, outer zones have more sectors/track than inner zones. § So we use average number of sectors/track when compu;ng capacity. Spindle … Carnegie Mellon 20 Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspec;ve, Third Edi;on Disk OperaCon (MulC-‐PlaXer View) Arm Read/write heads move in unison from cylinder to cylinder Spindle Carnegie Mellon 21 Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspec;ve, Third Edi;on Tracks divided into sectors Disk Structure -‐ top view of single plaXer Surface organized into tracks Carnegie Mellon 22 Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspec;ve, Third Edi;on Disk Access Head in position above a track Carnegie Mellon 25 Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspec;ve, Third Edi;on Disk Access – Read After BLUE read After reading blue sector Carnegie Mellon 26 Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspec;ve, Third Edi;on Disk Access – Read After BLUE read Red request scheduled next Carnegie Mellon 27 Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspec;ve, Third Edi;on Disk Access – Seek After BLUE read Seek for RED Seek to red’s track Carnegie Mellon 30 Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspec;ve, Third Edi;on Disk Access – Service Time Components After BLUE read Seek for RED Rotational latency After RED read Data transfer Seek RotaConal latency Data transfer Carnegie Mellon 31 Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspec;ve, Third Edi;on Disk Access Time ¢ Average Cme to access some target sector approximated by : § Taccess = Tavg seek + Tavg rota;on + Tavg transfer ¢ Seek Cme (Tavg seek) § Time to posi;on heads over cylinder containing target sector. § Typical Tavg seek is 3—9 ms ¢ RotaConal latency (Tavg rotaCon) § Time wai;ng for first bit of target sector to pass under r/w head. § Tavg rota;on = 1/2 x 1/RPMs x 60 sec/1 min § Typical Tavg rota;on = 7200 RPMs ¢ Transfer Cme (Tavg transfer) § Time to read the bits in the target sector. § Tavg transfer = 1/RPM x 1/(avg # sectors/track) x 60 secs/1 min. Carnegie Mellon 32 Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspec;ve, Third Edi;on Disk Access Time Example ¢ Given: § Rota;onal rate = 7,200 RPM § Average seek ;me = 9 ms. § Avg # sectors/track = 400. ¢ Derived: § Tavg rota;on = 1/2 x (60 secs/7200 RPM) x 1000 ms/sec = 4 ms. § Tavg transfer = 60/7200 RPM x 1/400 secs/track x 1000 ms/sec = 0.02 ms § Taccess = 9 ms + 4 ms + 0.02 ms ¢ Important points: § Access ;me dominated by seek ;me and rota;onal latency. § First bit in a sector is the most expensive, the rest are free. § SRAM access ;me is about 4 ns/doubleword, DRAM about 60 ns § Disk is about 40,000 ;mes slower than SRAM, § 2,500 ;mes slower then DRAM. Carnegie Mellon 35 Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspec;ve, Third Edi;on Reading a Disk Sector (1) Main memory ALU Register file CPU chip Disk controller Graphics adapter USB controller mouse keyboard Monitor Disk I/O bus Bus interface CPU initiates a disk read by writing a command, logical block number, and destination memory address to a port (address) associated with disk controller. Carnegie Mellon 36 Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspec;ve, Third Edi;on Reading a Disk Sector (2) Main memory ALU Register file CPU chip Disk controller Graphics adapter USB controller Mouse Keyboard Monitor Disk I/O bus Bus interface Disk controller reads the sector and performs a direct memory access (DMA) transfer into main memory. Carnegie Mellon 37 Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspec;ve, Third Edi;on Reading a Disk Sector (3) Main memory ALU Register file CPU chip Disk controller Graphics adapter USB controller Mouse Keyboard Monitor Disk I/O bus Bus interface When the DMA transfer completes, the disk controller notifies the CPU with an interrupt (i.e., asserts a special “interrupt” pin on the CPU) Carnegie Mellon 40 Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspec;ve, Third Edi;on SSD Tradeoffs vs RotaCng Disks ¢ Advantages § No moving parts à faster, less power, more rugged ¢ Disadvantages § Have the poten;al to wear out § Mi;gated by “wear leveling logic” in flash transla;on layer § E.g. Intel SSD 730 guarantees 128 petabyte (128 x 1015 bytes) of writes before they wear out § In 2015, about 30 ;mes more expensive per byte ¢ ApplicaCons § MP3 players, smart phones, laptops § Beginning to appear in desktops and servers Carnegie Mellon 41 Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspec;ve, Third Edi;on The CPU-‐Memory Gap The gap between DRAM, disk, and CPU speeds. 0.0 0.1 1.0 10.0 100.0 1,000.0 10,000.0 100,000.0 1,000,000.0 10,000,000.0 100,000,000.0 1985 1990 1995 2000 2003 2005 2010 2015 Ti m e (n s) Year Disk seek time SSD access time DRAM access time SRAM access time CPU cycle time Effective CPU cycle time DRAM CPU SSD Disk Carnegie Mellon 42 Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspec;ve, Third Edi;on Locality to the Rescue! The key to bridging this CPU-‐Memory gap is a fundamental property of computer programs known as locality Carnegie Mellon 45 Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspec;ve, Third Edi;on Locality Example ¢ Data references § Reference array elements in succession (stride-‐1 reference pacern). § Reference variable sum each itera;on. ¢ InstrucCon references § Reference instruc;ons in sequence. § Cycle through loop repeatedly. sum = 0; for (i = 0; i < n; i++) sum += a[i]; return sum; SpaCal locality Temporal locality SpaCal locality Temporal locality Carnegie Mellon 46 Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspec;ve, Third Edi;on QualitaCve EsCmates of Locality ¢ Claim: Being able to look at code and get a qualitaCve sense of its locality is a key skill for a professional programmer. ¢ QuesCon: Does this funcCon have good locality with respect to array a? int sum_array_rows(int a[M][N]) { int i, j, sum = 0; for (i = 0; i < M; i++) for (j = 0; j < N; j++) sum += a[i][j]; return sum; } Carnegie Mellon 47 Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspec;ve, Third Edi;on Locality Example ¢ QuesCon: Does this funcCon have good locality with respect to array a? int sum_array_cols(int a[M][N]) { int i, j, sum = 0; for (j = 0; j < N; j++) for (i = 0; i < M; i++) sum += a[i][j]; return sum; } Carnegie Mellon 50 Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspec;ve, Third Edi;on Today ¢ Storage technologies and trends ¢ Locality of reference ¢ Caching in the memory hierarchy Carnegie Mellon 51 Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspec;ve, Third Edi;on Example Memory Hierarchy Regs L1 cache (SRAM) Main memory (DRAM) Local secondary storage (local disks) Larger, slower, and cheaper (per byte) storage devices Remote secondary storage (e.g., Web servers) Local disks hold files retrieved from disks on remote servers L2 cache (SRAM) L1 cache holds cache lines retrieved from the L2 cache. CPU registers hold words retrieved from the L1 cache. L2 cache holds cache lines retrieved from L3 cache L0: L1: L2: L3: L4: L5: Smaller, faster, and costlier (per byte) storage devices L3 cache (SRAM) L3 cache holds cache lines retrieved from main memory. L6: Main memory holds disk blocks retrieved from local disks. Carnegie Mellon 52 Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspec;ve, Third Edi;on Caches ¢ Cache: A smaller, faster storage device that acts as a staging area for a subset of the data in a larger, slower device. ¢ Fundamental idea of a memory hierarchy: § For each k, the faster, smaller device at level k serves as a cache for the larger, slower device at level k+1. ¢ Why do memory hierarchies work? § Because of locality, programs tend to access the data at level k more oZen than they access the data at level k+1. § Thus, the storage at level k+1 can be slower, and thus larger and cheaper per bit. ¢ Big Idea: The memory hierarchy creates a large pool of storage that costs as much as the cheap storage near the boXom, but that serves data to programs at the rate of the fast storage near the top. Carnegie Mellon 55 Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspec;ve, Third Edi;on General Cache Concepts: Miss 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 8 9 14 3 Cache Memory Data in block b is needed Request: 12 Block b is not in cache: Miss! Block b is fetched from memory Request: 12 12 12 Block b is stored in cache • Placement policy: determines where b goes • Replacement policy: determines which block gets evicted (vic;m) Carnegie Mellon 56 Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspec;ve, Third Edi;on General Caching Concepts: Types of Cache Misses ¢ Cold (compulsory) miss § Cold misses occur because the cache is empty. ¢ Conflict miss § Most caches limit blocks at level k+1 to a small subset (some;mes a singleton) of the block posi;ons at level k. § E.g. Block i at level k+1 must be placed in block (i mod 4) at level k. § Conflict misses occur when the level k cache is large enough, but mul;ple data objects all map to the same level k block. § E.g. Referencing blocks 0, 8, 0, 8, 0, 8, ... would miss every ;me. ¢ Capacity miss § Occurs when the set of ac;ve cache blocks (working set) is larger than the cache. Carnegie Mellon 57 Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspec;ve, Third Edi;on Examples of Caching in the Mem. Hierarchy Hardware MMU 0 On-‐Chip TLB Address translaCons TLB Web browser 10,000,000 Local disk Web pages Browser cache Web cache Network buffer cache Buffer cache Virtual Memory L2 cache L1 cache Registers Cache Type Web pages Parts of files Parts of files 4-‐KB pages 64-‐byte blocks 64-‐byte blocks 4-‐8 bytes words What is Cached? Web proxy server 1,000,000,000 Remote server disks OS 100 Main memory Hardware 4 On-‐Chip L1 Hardware 10 On-‐Chip L2 NFS client 10,000,000 Local disk Hardware + OS 100 Main memory Compiler 0 CPU core Managed By Latency (cycles) Where is it Cached? Disk cache Disk sectors Disk controller 100,000 Disk firmware Carnegie Mellon 60 Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspec;ve, Third Edi;on ConvenConal DRAM OrganizaCon ¢ d x w DRAM: § dw total bits organized as d supercells of size w bits cols rows 0 1 2 3 0 1 2 3 Internal row buffer 16 x 8 DRAM chip addr data supercell (2,1) 2 bits / 8 bits / Memory controller (to/from CPU) Carnegie Mellon 61 Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspec;ve, Third Edi;on Reading DRAM Supercell (2,1) Step 1(a): Row access strobe (RAS) selects row 2. Step 1(b): Row 2 copied from DRAM array to row buffer. Cols Rows RAS = 2 0 1 2 3 0 1 2 Internal row buffer 16 x 8 DRAM chip 3 addr data 2 / 8 / Memory controller Carnegie Mellon 62 Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspec;ve, Third Edi;on Reading DRAM Supercell (2,1) Step 2(a): Column access strobe (CAS) selects column 1. Step 2(b): Supercell (2,1) copied from buffer to data lines, and eventually back to the CPU. Cols Rows 0 1 2 3 0 1 2 3 Internal row buffer 16 x 8 DRAM chip CAS = 1 addr data 2 / 8 / Memory controller supercell (2,1) supercell (2,1) To CPU Carnegie Mellon 65 Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspec;ve, Third Edi;on Metric 1985 1990 1995 2000 2005 2010 2015 2015:1985 $/MB 880 100 30 1 0.1 0.06 0.02 44,000 access (ns) 200 100 70 60 50 40 20 10 typical size (MB) 0.256 4 16 64 2,000 8,000 16.000 62,500 Storage Trends DRAM SRAM Metric 1985 1990 1995 2000 2005 2010 2015 2015:1985 $/GB 100,000 8,000 300 10 5 0.3 0.03 3,333,333 access (ms) 75 28 10 8 5 3 3 25 typical size (GB) 0.01 0.16 1 20 160 1,500 3,000 300,000 Disk Metric 1985 1990 1995 2000 2005 2010 2015 2015:1985 $/MB 2,900 320 256 100 75 60 320 116 access (ns) 150 35 15 3 2 1.5 200 115 Carnegie Mellon 66 Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspec;ve, Third Edi;on CPU Clock Rates 1985 1990 1995 2003 2005 2010 2015 2015:1985 CPU 80286 80386 Pentium P-4 Core 2 Core i7(n) Core i7(h) Clock rate (MHz) 6 20 150 3,300 2,000 2,500 3,000 500 Cycle time (ns) 166 50 6 0.30 0.50 0.4 0.33 500 Cores 1 1 1 1 2 4 4 4 Effective cycle 166 50 6 0.30 0.25 0.10 0.08 2,075 time (ns) InflecCon point in computer history when designers hit the “Power Wall” (n) Nehalem processor (h) Haswell processor