Data Storage the CPU-Memory

Total Page：16

File Type：pdf, Size：1020Kb

Data Storage •Disks • Hard disk (HDD) • Solid state drive (SSD) •Random Access Memory • Dynamic RAM (DRAM) • Static RAM (SRAM) •Registers • %rax, %rbx, ... Sean Barker 1 The CPU-Memory Gap 100,000,000.0 10,000,000.0 Disk 1,000,000.0 100,000.0 SSD Disk seek time 10,000.0 SSD access time 1,000.0 DRAM access time Time (ns) Time 100.0 DRAM SRAM access time CPU cycle time 10.0 Effective CPU cycle time 1.0 0.1 CPU 0.0 1985 1990 1995 2000 2003 2005 2010 2015 Year Sean Barker 2 Caching Smaller, faster, more expensive Cache 8 4 9 10 14 3 memory caches a subset of the blocks Data is copied in block-sized 10 4 transfer units Larger, slower, cheaper memory Memory 0 1 2 3 viewed as par@@oned into “blocks” 4 5 6 7 8 9 10 11 12 13 14 15 Sean Barker 3 Cache Hit Request: 14 Cache 8 9 14 3 Memory 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Sean Barker 4 Cache Miss Request: 12 Cache 8 12 9 14 3 12 Request: 12 Memory 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Sean Barker 5 Locality ¢ Temporal locality: ¢ Spa0al locality: Sean Barker 6 Locality Example (1) sum = 0; for (i = 0; i < n; i++) sum += a[i]; return sum; Sean Barker 7 Locality Example (2) 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; } Sean Barker 8 Locality Example (3) 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; } Sean Barker 9 The Memory Hierarchy The Memory Hierarchy Smaller On 1 cycle to access CPU Chip Registers Faster Storage Costlier instrs can L1, L2 per byte directly Cache(s) ~10’s of cycles to access access (SRAM) Main memory ~100 cycles to access (DRAM) Larger Slower Flash SSD / Local network ~100 M cycles to access Cheaper Local secondary storage (disk) per byte slower Remote secondary storage than local (tapes, Web servers / Internet) disk to access Sean Barker 10.

Copy Link

Recommended publications

Managing the Memory Hierarchy
Managing the Memory Hierarchy Jeffrey S. Vetter Sparsh Mittal, Joel Denny, Seyong Lee Presented to SOS20 Asheville 24 Mar 2016 ORNL is managed by UT-Battelle for the US Department of Energy http://ft.ornl.gov [email protected] Exascale architecture targets circa 2009 2009 Exascale Challenges Workshop in San Diego Attendees envisioned two possible architectural swim lanes: 1. Homogeneous many-core thin-node system 2. Heterogeneous (accelerator + CPU) fat-node system System attributes 2009 “Pre-Exascale” “Exascale” System peak 2 PF 100-200 PF/s 1 Exaflop/s Power 6 MW 15 MW 20 MW System memory 0.3 PB 5 PB 32–64 PB Storage 15 PB 150 PB 500 PB Node performance 125 GF 0.5 TF 7 TF 1 TF 10 TF Node memory BW 25 GB/s 0.1 TB/s 1 TB/s 0.4 TB/s 4 TB/s Node concurrency 12 O(100) O(1,000) O(1,000) O(10,000) System size (nodes) 18,700 500,000 50,000 1,000,000 100,000 Node interconnect BW 1.5 GB/s 150 GB/s 1 TB/s 250 GB/s 2 TB/s IO Bandwidth 0.2 TB/s 10 TB/s 30-60 TB/s MTTI day O(1 day) O(0.1 day) 2 Memory Systems • Multimode memories – Fused, shared memory – Scratchpads – Write through, write back, etc – Virtual v. Physical, paging strategies – Consistency and coherence protocols • 2.5D, 3D Stacking • HMC, HBM/2/3, LPDDR4, GDDR5, WIDEIO2, etc https://www.micron.com/~/media/track-2-images/content-images/content_image_hmc.jpg?la=en • New devices (ReRAM, PCRAM, Xpoint) J.S.
[Show full text]
Embedded DRAM
Embedded DRAM Raviprasad Kuloor Semiconductor Research and Development Centre, Bangalore IBM Systems and Technology Group DRAM Topics Introduction to memory DRAM basics and bitcell array eDRAM operational details (case study) Noise concerns Wordline driver (WLDRV) and level translators (LT) Challenges in eDRAM Understanding Timing diagram – An example References Slide 1 Acknowledgement • John Barth, IBM SRDC for most of the slides content • Madabusi Govindarajan • Subramanian S. Iyer • Many Others Slide 2 Topics Introduction to memory DRAM basics and bitcell array eDRAM operational details (case study) Noise concerns Wordline driver (WLDRV) and level translators (LT) Challenges in eDRAM Understanding Timing diagram – An example Slide 3 Memory Classification revisited Slide 4 Motivation for a memory hierarchy – infinite memory Memory store Processor Infinitely fast Infinitely large Cycles per Instruction Number of processor clock cycles (CPI) = required per instruction CPI[ ∞ cache] Finite memory speed Memory store Processor Finite speed Infinite size CPI = CPI[∞ cache] + FCP Finite cache penalty Locality of reference – spatial and temporal Temporal If you access something now you’ll need it again soon e.g: Loops Spatial If you accessed something you’ll also need its neighbor e.g: Arrays Exploit this to divide memory into hierarchy Hit L2 L1 (Slow) Processor Miss (Fast) Hit Register Cache size impacts cycles-per-instruction Access rate reduces Slower memory is sufficient Cache size impacts cycles-per-instruction For a 5GHz
[Show full text]
Cache-Aware Roofline Model: Upgrading the Loft
1 Cache-aware Roofline model: Upgrading the loft Aleksandar Ilic, Frederico Pratas, and Leonel Sousa INESC-ID/IST, Technical University of Lisbon, Portugal ilic,fcpp,las @inesc-id.pt f g Abstract—The Roofline model graphically represents the attainable upper bound performance of a computer architecture. This paper analyzes the original Roofline model and proposes a novel approach to provide a more insightful performance modeling of modern architectures by introducing cache-awareness, thus significantly improving the guidelines for application optimization. The proposed model was experimentally verified for different architectures by taking advantage of built-in hardware counters with a curve fitness above 90%. Index Terms—Multicore computer architectures, Performance modeling, Application optimization F 1 INTRODUCTION in Tab. 1. The horizontal part of the Roofline model cor- Driven by the increasing complexity of modern applica- responds to the compute bound region where the Fp can tions, microprocessors provide a huge diversity of compu- be achieved. The slanted part of the model represents the tational characteristics and capabilities. While this diversity memory bound region where the performance is limited is important to fulfill the existing computational needs, it by BD. The ridge point, where the two lines meet, marks imposes challenges to fully exploit architectures’ potential. the minimum I required to achieve Fp. In general, the A model that provides insights into the system performance original Roofline modeling concept [10] ties the Fp and the capabilities is a valuable tool to assist in the development theoretical bandwidth of a single memory level, with the I and optimization of applications and future architectures.
[Show full text]
ARM-Architecture Simulator Archulator
Embedded ECE Lab Arm Architecture Simulator University of Arizona ARM-Architecture Simulator Archulator Contents Purpose ............................................................................................................................................ 2 Description ...................................................................................................................................... 2 Block Diagram ............................................................................................................................ 2 Component Description .............................................................................................................. 3 Buses ..................................................................................................................................................... 3 µP .......................................................................................................................................................... 3 Caches ................................................................................................................................................... 3 Memory ................................................................................................................................................. 3 CoProcessor .......................................................................................................................................... 3 Using the Archulator ......................................................................................................................
[Show full text]
Multi-Tier Caching Technology™
Multi-Tier Caching Technology™ Technology Paper Authored by: How Layering an Application’s Cache Improves Performance Modern data storage needs go far beyond just computing. From creative professional environments to desktop systems, Seagate provides solutions for almost any application that requires large volumes of storage. As a leader in NAND, hybrid, SMR and conventional magnetic recording technologies, Seagate® applies different levels of caching and media optimization to benefit performance and capacity. Multi-Tier Caching (MTC) Technology brings the highest performance and areal density to a multitude of applications. Each Seagate product is uniquely tailored to meet the performance requirements of a specific use case with the right memory, NAND, and media type and size. This paper explains how MTC Technology works to optimize hard drive performance. MTC Technology: Key Advantages Capacity requirements can vary greatly from business to business. While the fastest performance can be achieved using Dynamic Random Access Memory (DRAM) cache, the data in the DRAM is not persistent through power cycles and DRAM is very expensive compared to other media. NAND flash data survives through power cycles but it is still very expensive compared to a magnetic storage medium. Magnetic storage media cache offers good performance at a very low cost. On the downside, media cache takes away overall disk drive capacity from PMR or SMR main store media. MTC Technology solves this dilemma by using these diverse media components in combination to offer different levels of performance and capacity at varying price points. By carefully tuning firmware with appropriate cache types and sizes, the end user can experience excellent overall system performance.
[Show full text]
CUDA 11 and A100 - WHAT’S NEW? Markus Hrywniak, 23Rd June 2020 TOPICS for TODAY
CUDA 11 AND A100 - WHAT’S NEW? Markus Hrywniak, 23rd June 2020 TOPICS FOR TODAY Ampere architecture – A100, powering DGX–A100, HGX-A100... and soon, FZ Jülich‘s JUWELS Booster New CUDA 11 Toolkit release Overview of features Talk next week: Third generation Tensor Cores GTC talks go into much more details. See references! 2 HGX-A100 4-GPU HGX-A100 8-GPU • 4 A100 with NVLINK • 8 A100 with NVSwitch 3 HIERARCHY OF SCALES Multi-System Rack Multi-GPU System Multi-SM GPU Multi-Core SM Unlimited Scale 8 GPUs 108 Multiprocessors 2048 threads 4 AMDAHL’S LAW serial section parallel section serial section Amdahl’s Law parallel section Shortest possible runtime is sum of serial section times Time saved serial section Some Parallelism Increased Parallelism Infinite Parallelism Program time = Parallel sections take less time Parallel sections take no time sum(serial times + parallel times) Serial sections take same time Serial sections take same time 5 OVERCOMING AMDAHL: ASYNCHRONY & LATENCY serial section parallel section serial section Split up serial & parallel components parallel section serial section Some Parallelism Task Parallelism Infinite Parallelism Program time = Parallel sections overlap with serial sections Parallel sections take no time sum(serial times + parallel times) Serial sections take same time 6 OVERCOMING AMDAHL: ASYNCHRONY & LATENCY CUDA Concurrency Mechanisms At Every Scope CUDA Kernel Threads, Warps, Blocks, Barriers Application CUDA Streams, CUDA Graphs Node Multi-Process Service, GPU-Direct System NCCL, CUDA-Aware MPI, NVSHMEM 7 OVERCOMING AMDAHL: ASYNCHRONY & LATENCY Execution Overheads Non-productive latencies (waste) Operation Latency Network latencies Memory read/write File I/O ..
[Show full text]
The Memory Hierarchy
Chapter 6 The Memory Hierarchy To this point in our study of systems, we have relied on a simple model of a computer system as a CPU that executes instructions and a memory system that holds instructions and data for the CPU. In our simple model, the memory system is a linear array of bytes, and the CPU can access each memory location in a constant amount of time. While this is an effective model as far as it goes, it does not reﬂect the way that modern systems really work. In practice, a memory system is a hierarchy of storage devices with different capacities, costs, and access times. CPU registers hold the most frequently used data. Small, fast cache memories nearby the CPU act as staging areas for a subset of the data and instructions stored in the relatively slow main memory. The main memory stages data stored on large, slow disks, which in turn often serve as staging areas for data stored on the disks or tapes of other machines connected by networks. Memory hierarchies work because well-written programs tend to access the storage at any particular level more frequently than they access the storage at the next lower level. So the storage at the next level can be slower, and thus larger and cheaper per bit. The overall effect is a large pool of memory that costs as much as the cheap storage near the bottom of the hierarchy, but that serves data to programs at the rate of the fast storage near the top of the hierarchy.
[Show full text]
Lecture 10: Memory Hierarchy -- Memory Technology and Principal of Locality
Lecture 10: Memory Hierarchy -- Memory Technology and Principal of Locality CSCE 513 Computer Architecture Department of Computer Science and Engineering Yonghong Yan [email protected] https://passlab.github.io/CSCE513 1 Topics for Memory Hierarchy • Memory Technology and Principal of Locality – CAQA: 2.1, 2.2, B.1 – COD: 5.1, 5.2 • Cache Organization and Performance – CAQA: B.1, B.2 – COD: 5.2, 5.3 • Cache Optimization – 6 Basic Cache Optimization Techniques • CAQA: B.3 – 10 Advanced Optimization Techniques • CAQA: 2.3 • Virtual Memory and Virtual Machine – CAQA: B.4, 2.4; COD: 5.6, 5.7 – Skip for this course 2 The Big Picture: Where are We Now? Processor Input Control Memory Datapath Output • Memory system – Supplying data on time for computation (speed) – Large enough to hold everything needed (capacity) 3 Overview • Programmers want unlimited amounts of memory with low latency • Fast memory technology is more expensive per bit than slower memory • Solution: organize memory system into a hierarchy – Entire addressable memory space available in largest, slowest memory – Incrementally smaller and faster memories, each containing a subset of the memory below it, proceed in steps up toward the processor • Temporal and spatial locality insures that nearly all references can be found in smaller memories – Gives the allusion of a large, fast memory being presented to the processor 4 Memory Hierarchy Memory Hierarchy 5 Memory Technology • Random Access: access time is the same for all locations • DRAM: Dynamic Random Access Memory – High density, low power, cheap, slow – Dynamic: need to be “refreshed” regularly – 50ns – 70ns, $20 – $75 per GB • SRAM: Static Random Access Memory – Low density, high power, expensive, fast – Static: content will last “forever”(until lose power) – 0.5ns – 2.5ns, $2000 – $5000 per GB Ideal memory: • Magnetic disk • Access time of SRAM – 5ms – 20ms, $0.20 – $2 per GB • Capacity and cost/GB of disk 6 Static RAM (SRAM) 6-Transistor Cell – 1 Bit 6-Transistor SRAM Cell word word 0 1 (row select) 0 1 bit bit bit bit • Write: 1.
[Show full text]
Chapter 2: Memory Hierarchy Design
Computer Architecture A Quantitative Approach, Fifth Edition Chapter 2 Memory Hierarchy Design Copyright © 2012, Elsevier Inc. All rights reserved. 1 Contents 1. Memory hierarchy 1. Basic concepts 2. Design techniques 2. Caches 1. Types of caches: Fully associative, Direct mapped, Set associative 2. Ten optimization techniques 3. Main memory 1. Memory technology 2. Memory optimization 3. Power consumption 4. Memory hierarchy case studies: Opteron, Pentium, i7. 5. Virtual memory 6. Problem solving dcm 2 Introduction Introduction Programmers want very large memory with low latency Fast memory technology is more expensive per bit than slower memory Solution: organize memory system into a hierarchy Entire addressable memory space available in largest, slowest memory Incrementally smaller and faster memories, each containing a subset of the memory below it, proceed in steps up toward the processor Temporal and spatial locality insures that nearly all references can be found in smaller memories Gives the allusion of a large, fast memory being presented to the processor Copyright © 2012, Elsevier Inc. All rights reserved. 3 Memory hierarchy Processor L1 Cache L2 Cache L3 Cache Main Memory Latency Hard Drive or Flash Capacity (KB, MB, GB, TB) Copyright © 2012, Elsevier Inc. All rights reserved. 4 PROCESSOR L1: I-Cache D-Cache I-Cache instruction cache D-Cache data cache U-Cache unified cache L2: U-Cache Different functional units fetch information from I-cache and D-cache: decoder and L3: U-Cache scheduler operate with I- cache, but integer execution unit and floating-point unit Main: Main Memory communicate with D-cache. Copyright © 2012, Elsevier Inc. All rights reserved.
[Show full text]
A Characterization of Processor Performance in the VAX-1 L/780
A Characterization of Processor Performance in the VAX-1 l/780 Joel S. Emer Douglas W. Clark Digital Equipment Corp. Digital Equipment Corp. 77 Reed Road 295 Foster Street Hudson, MA 01749 Littleton, MA 01460 ABSTRACT effect of many architectural and implementation features. This paper reports the results of a study of VAX- llR80 processor performance using a novel hardware Prior related work includes studies of opcode monitoring technique. A micro-PC histogram frequency and other features of instruction- monitor was buiit for these measurements. It kee s a processing [lo. 11,15,161; some studies report timing count of the number of microcode cycles execute z( at Information as well [l, 4,121. each microcode location. Measurement ex eriments were performed on live timesharing wor i loads as After describing our methods and workloads in well as on synthetic workloads of several types. The Section 2, we will re ort the frequencies of various histogram counts allow the calculation of the processor events in 5 ections 3 and 4. Section 5 frequency of various architectural events, such as the resents the complete, detailed timing results, and frequency of different types of opcodes and operand !!Iection 6 concludes the paper. specifiers, as well as the frequency of some im lementation-s ecific events, such as translation bu h er misses. ?phe measurement technique also yields the amount of processing time spent, in various 2. DEFINITIONS AND METHODS activities, such as ordinary microcode computation, memory management, and processor stalls of 2.1 VAX-l l/780 Structure different kinds. This paper reports in detail the amount of time the “average’ VAX instruction The llf780 processor is composed of two major spends in these activities.
[Show full text]
Instruction Set Innovations for Convey's HC-1 Computer
Instruction Set Innovations for Convey's HC-1 Computer THE WORLD’S FIRST HYBRID-CORE COMPUTER. Hot Chips Conference 2009 [email protected] Introduction to Convey Computer • Company Status – Second round venture based startup company – Product beta systems are at customer sites – Currently staffing at 36 people – Located in Richardson, Texas • Investors – Four Venture Capital Investors • Interwest Partners (Menlo Park) • CenterPoint Ventures (Dallas) • Rho Ventures (New York) • Braemar Energy Ventures (Boston) – Two Industry Investors • Intel Capital • Xilinx Presentation Outline • Overview of HC-1 Computer • Instruction Set Innovations • Application Examples Page 3 Hot Chips Conference 2009 What is a Hybrid-Core Computer ? A hybrid-core computer improves application performance by combining an x86 processor with hardware that implements application-specific instructions. ANSI Standard Applications C/C++/Fortran Convey Compilers x86 Coprocessor Instructions Instructions Intel® Processor Hybrid-Core Coprocessor Oil & Gas& Oil Financial Sciences Custom CAE Application-Specific Personalities Cache-coherent shared virtual memory Page 4 Hot Chips Conference 2009 What Is a Personality? • A personality is a reloadable set of instructions that augment x86 application the x86 instruction set Processor specific – Applicable to a class of applications instructions or specific to a particular code • Each personality is a set of files that includes: – The bits loaded into the Coprocessor – Information used by the Convey compiler • List of
[Show full text]
Memory Hierarchy Summary
Carnegie Mellon Carnegie Mellon Today DRAM as building block for main memory The Memory Hierarchy Locality of reference Caching in the memory hierarchy Storage technologies and trends 15‐213 / 18‐213: Introduction to Computer Systems 10th Lecture, Feb 14, 2013 Instructors: Seth Copen Goldstein, Anthony Rowe, Greg Kesden 1 2 Carnegie Mellon Carnegie Mellon Byte‐Oriented Memory Organization Simple Memory Addressing Modes Normal (R) Mem[Reg[R]] • • • . Register R specifies memory address . Aha! Pointer dereferencing in C Programs refer to data by address movl (%ecx),%eax . Conceptually, envision it as a very large array of bytes . In reality, it’s not, but can think of it that way . An address is like an index into that array Displacement D(R) Mem[Reg[R]+D] . and, a pointer variable stores an address . Register R specifies start of memory region . Constant displacement D specifies offset Note: system provides private address spaces to each “process” . Think of a process as a program being executed movl 8(%ebp),%edx . So, a program can clobber its own data, but not that of others nd th From 2 lecture 3 From 5 lecture 4 Carnegie Mellon Carnegie Mellon Traditional Bus Structure Connecting Memory Read Transaction (1) CPU and Memory CPU places address A on the memory bus. A bus is a collection of parallel wires that carry address, data, and control signals. Buses are typically shared by multiple devices. Register file Load operation: movl A, %eax ALU %eax CPU chip Main memory Register file I/O bridge A 0 Bus interface ALU x A System bus Memory bus I/O Main Bus interface bridge memory 5 6 Carnegie Mellon Carnegie Mellon Memory Read Transaction (2) Memory Read Transaction (3) Main memory reads A from the memory bus, retrieves CPU read word xfrom the bus and copies it into register word x, and places it on the bus.
[Show full text]