Hardware. THE MOVES INSTITUTE Hardware So you want to build a cluster. What do you need to buy? Remember the definition of a beowulf cluster: Commodity.

Slides:



Advertisements
Similar presentations
Section 6.2. Record data by magnetizing the binary code on the surface of a disk. Data area is reusable Allows for both sequential and direct access file.
Advertisements

Introduction to Computer Hardware and Software. Definition of a Computer “A computer is an electronic device, operating under the control of instructions.
Terms 4 Definitions and Questions. Motherboard The main board of a computer, usually containing the circuitry for the central processing unit, keyboard,
♦ Commodity processor with commodity inter- processor connection Clusters Pentium, Itanium, Opteron, Alpha GigE, Infiniband, Myrinet, Quadrics, SCI NEC.
Beowulf Supercomputer System Lee, Jung won CS843.
BY TONY JIA Mother Board and Buses. What is a Mother Board? The motherboard is the largest piece of internal hardware. All of the other internal.
Lecture 12 Page 1 CS 111 Online Devices and Device Drivers CS 111 On-Line MS Program Operating Systems Peter Reiher.
Types of Parallel Computers
Linux Clustering A way to supercomputing. What is Cluster? A group of individual computers bundled together using hardware and software in order to make.
Presented by: Yash Gurung, ICFAI UNIVERSITY.Sikkim BUILDING of 3 R'sCLUSTER PARALLEL COMPUTER.
Novell Server Linux vs. windows server 2008 By: Gabe Miller.
Computer Terminology … Remember: Knowledge is Power!
Group 11 Pekka Nikula Ossi Hämäläinen Introduction to Parallel Computing Kentucky Linux Athlon Testbed 2
Threads CS 416: Operating Systems Design, Spring 2001 Department of Computer Science Rutgers University
WHAT IS A COMPUTER? BY JACK SUMMERS. WHAT IS A COMPUTER? A computer basically a set of different components that when put together in the correct way.
Hardware and Software Basics. Computer Hardware  Central Processing Unit - also called “The Chip”, a CPU, a processor, or a microprocessor  Memory (RAM)
Virtual Network Servers. What is a Server? 1. A software application that provides a specific one or more services to other computers  Example: Apache.
CPP Staff - 30 CPP Staff - 30 FCIPT Staff - 35 IPR Staff IPR Staff ITER-India Staff ITER-India Staff Research Areas: 1.Studies.
1b.1 Types of Parallel Computers Two principal approaches: Shared memory multiprocessor Distributed memory multicomputer ITCS 4/5145 Parallel Programming,
THE CPU Cpu brands AMD cpu Intel cpu By Nathan Ferguson.
Basics of Operating Systems March 4, 2001 Adapted from Operating Systems Lecture Notes, Copyright 1997 Martin C. Rinard.
Real Parallel Computers. Modular data centers Background Information Recent trends in the marketplace of high performance computing Strohmaier, Dongarra,
Cluster Computers. Introduction Cluster computing –Standard PCs or workstations connected by a fast network –Good price/performance ratio –Exploit existing.
5.3 HS23 Blade Server. The HS23 blade server is a dual CPU socket blade running Intel´s new Xeon® processor, the E5-2600, and is the first IBM BladeCenter.
Recent experience in buying and configuring a cluster John Matrow, Director High Performance Computing Center Wichita State University.
Cluster computing facility for CMS simulation work at NPD-BARC Raman Sehgal.
Day 10 Hardware Fault Tolerance RAID. High availability All servers should be on UPSs –2 Types Smart UPS –Serial cable connects from UPS to computer.
A+ Guide to Hardware: Managing, Maintaining, and Troubleshooting, Sixth Edition Chapter 9, Part 11 Satisfying Customer Needs.
… when you will open a computer We hope you will not look like …
Introduction to Computers Personal Computing 10. What is a computer? Electronic device Performs instructions in a program Performs four functions –Accepts.
Serial vs.Parallel Computing Scalable Perf. vs. Availability
B.A. (Mahayana Studies) Introduction to Computer Science November March The Motherboard A look at the brains of the computer, the.
DIY: Your First VMware Server. Introduction to ESXi, VMWare's free virtualization Operating System.
Enabling Technologies for Distributed and Cloud Computing Dr. Sanjay P. Ahuja, Ph.D FIS Distinguished Professor of Computer Science School of.
Building a PC Tim Wells. What You Need Required:  Case  Power Supply  Motherboard  Processor & CPU Fan  RAM  Graphics Card  Hard Drive  Keyboard.
 Design model for a computer  Named after John von Neuman  Instructions that tell the computer what to do are stored in memory  Stored program Memory.
Translate the following message:
By: Joshua chambers INTRODUCTIONINTRODUCTION What your computer can do depends upon two things: the hardware your computer has, and the software that.
How Computers Work The Four Basic Operations The Boot Process
1b.1 Types of Parallel Computers Two principal approaches: Shared memory multiprocessor Distributed memory multicomputer ITCS 4/5145 Parallel Programming,
A mother board is the main circuit board for the computer system (hence the name mother board ) The mother board holds all of the other components of a.
Multi-core Programming Introduction Topics. Topics General Ideas Moore’s Law Amdahl's Law Processes and Threads Concurrency vs. Parallelism.
By: Dorian Gobert. In 1998, the Gateway G6-450 was "top of the line", the fastest computer Gateway offered. Compared it to the eMachines ET in.
IT253: Computer Organization
Basic Computer Components INTRODUCTION TO COMPUTER.
The Guts. CPU CPU Socket The CPU is generally a 2 inch ceramic square with a silicon chip located inside. The chip usually about the size of a thumbnail.
Beowulf – Cluster Nodes & Networking Hardware Garrison Vaughan.
Enabling Technologies for Distributed Computing Dr. Sanjay P. Ahuja, Ph.D. Fidelity National Financial Distinguished Professor of CIS School of Computing,
COMP381 by M. Hamdi 1 Clusters: Networks of WS/PC.
Parts of the computer Deandre Haynes. The Case The Case This Case is the "box" or "chassis" that holds and encloses the many parts of your computer. Its.
Learning Computer Concepts Lesson 1: Defining a Computer Lesson Objectives List types of computers Identify benefits of computers Describe how a computer.
Computer Hardware & Processing Inside the Box CSC September 16, 2010.
Cluster Computers. Introduction Cluster computing –Standard PCs or workstations connected by a fast network –Good price/performance ratio –Exploit existing.
Background Computer System Architectures Computer System Software.
By Harshal Ghule Guided by Mrs. Anita Mahajan G.H.Raisoni Institute Of Engineering And Technology.
Copyright © 2006 by The McGraw-Hill Companies, Inc. All rights reserved. McGraw-Hill Technology Education Chapter 1 Looking Inside the Computer System.
Hardware Architecture
Constructing a system with multiple computers or processors 1 ITCS 4/5145 Parallel Programming, UNC-Charlotte, B. Wilkinson. Jan 13, 2016.
Chapter 1: Introduction to the Personal Computer
Computer Hardware.
CS 6560: Operating Systems Design
Constructing a system with multiple computers or processors
Drill Translate the following message:
Instructor Materials Chapter 1: Introduction to the Personal Computer
Constructing a system with multiple computers or processors
Constructing a system with multiple computers or processors
Constructing a system with multiple computers or processors
Overview 1. Inside a PC 2. The Motherboard 3. RAM the 'brains' 4. ROM
Types of Parallel Computers
Cluster Computers.
Presentation transcript:

Hardware

THE MOVES INSTITUTE Hardware So you want to build a cluster. What do you need to buy? Remember the definition of a beowulf cluster: Commodity machines Private cluster network Open source software Two of these are hardware-related

THE MOVES INSTITUTE Racks You need some place to keep all the hardware organized. Racks are the usual solution. The things you put into racks are measured in something called “U”s. The standard rack is 42 U’s tall--A U is a measure of vertical height, about 1.75”. CPUs are often one or two U’s, ie they take up one or two slots on the rack.

THE MOVES INSTITUTE Racks It’s easy to get a cable mess in the back of the rack. Aside from management, a cable mess can affect air flow. Be consistent about routing power and network cables. Color-coded cables are an excellent idea; for example, make all private network cables one color, public network cables another Use zip ties to group cables Make sure you have enough space for air flow (typically front to back)

THE MOVES INSTITUTE KVM Keyboard-Video-Mouse. While usually you want to remotely manage the cluster, you’ll often wind up attaching a monitor to a box to watch it boot or troubleshoot. It’s good to have a lightweight flat panel display, keyboard, and mouse you can attach as needed

THE MOVES INSTITUTE CPU: 64 vs. 32 Bit Isn’t 64 bits twice as good as 32 bits? Not really. The 64 bits usually refers to the amount of memory that can be addressed. With 32 bits you can ‘count’ up to 4 billion, which usually means you can’t have more than 4 GB of memory in one process. Because of the way VM is set up, 2 GB is a more typical maximum address space for 32 bit processes 64 bits allows 4.5 Petabytes of memory address This can be important if you have an application that breaks the 2 GB address space barrier This is plausible now that many machines can be configured with more than 4 GB of real memory

THE MOVES INSTITUTE 64 vs. 32 bit 64 bit CPUs have bigger registers (high speed holding areas) but this often isn’t all that big of a deal. Almost every 64 bit CPU has a compatibility mode that allows 32 bit applications to be run To run in 64 bit mode your application will need to be compiled for 64 bits and link to 64 bit libraries.

THE MOVES INSTITUTE 64 bit: The Dark Side A pointer is just a piece of data that holds the address of something in memory int count = 100; int *aPtr = &count; GB 0x2248 aPtr

THE MOVES INSTITUTE 64 bit: The Dark Side Since a pointer refers to a place in memory, it needs to be big enough to reference any place in the memory range. For 32 bit applications that is 4 bytes. For 64 bit applications that is 8 bytes, twice as big. That means code that contains 64 bit pointers will take up more space than code with 32 bit pointers. If you have a cache with a fixed size you can fit less 64 bit code into it p1p2 p3p4 p5p6 24 byte cache can hold six 32 bit pointers p1 p2 p3 24 byte cache can hold three 64 bit pointers

THE MOVES INSTITUTE 64 bit: Dark Side The extra space that 64 bit code takes up in caches can reduce performance by increasing cache misses. For this reason you probably shouldn’t compile to 64 bits unless your application requires a large address space (In fact under OS X the GUI libraries are all 32 bit, even though the underlying hardware may be 64 bit. Compiling them to 64 bit would have reduced performance.)

THE MOVES INSTITUTE Dual Core CPU designers have been running up against a wall lately; they can’t increase clock speed as much as in the past As a result they are adding more silicon, in the case of dual-core CPUs replicating two or more CPUs and cache on one chip This doesn’t increase sequential speed, but can increase speed for multiple process or multiple thread programs

THE MOVES INSTITUTE CPUs This is a major religious issue. The major contenders in the HPC Beowulf cluster space are AMD Opteron (64/32 bit) Intel Itanium (64/32 bit) Intel Xeon MP (Dual-core, EM64T) Intel Pentium M (32 bit, low power) PowerPC G5/970 (64 bit, OS X usually)

THE MOVES INSTITUTE CPUs The Opteron is a strong choice. Good price/performance, good power consumption, good SMP capabilities But some vendors (famously including Dell) don’t sell Opterons

THE MOVES INSTITUTE SMP Each node in the cluster can be an SMP machine, for example a dual-processor 1U box. If the CPU is dual core this can give 4 cores per box Four and eight CPU SMP boxes are also available. More cores per CPU are likely in the future.

THE MOVES INSTITUTE Blades Lots of people want extremely dense installations, with the most possible CPUs per square foot. Blades are essentially hosts on plug-in cards. They’re inserted into a blade chassis.

THE MOVES INSTITUTE Blades The blades have an interconnect to the chassis and can share some resources such as the power supply, CD and floppy, network cables, etc. They may implement features such as hot swap The downside is that they can be somewhat more expensive than plain 1U rackmounts, may be more proprietary, and are somewhat less flexible. Also, while they take up less floor space, they still generate a similar amount of heat IBM is a major blade vendor

THE MOVES INSTITUTE Memory How much memory do you need on a compute node? That depends on your application and current computer part economics. Roughly 1/2 GB per GFLOP of speed seems to do OK. Opterons are at this writing roughly 3 GFLOP/processor, so about 3-4 GB per dual processor box. This may change with dual core CPUs. You don’t want your application to page-fault. If you’re running 64 bit that probably means you need an address space over 4 GB, so you may well need 4-8 GB or more of physical memory

THE MOVES INSTITUTE Virtual Memory Virtual Process Address Space (Often 4 GB) Process Working Set in Physical Memory Max size = Physical memory Size Page faults happen when a page isn’t in working memory and has to be retrieved from disk. If you can fit the entire process into physical memory you can avoid page faults and speed up the process.

THE MOVES INSTITUTE Disk There are several schools of though about disk on compute nodes: Leave them diskless. - Less heat, less to fail, single image on server Put a big enough disk in to swap locally - Don’t have to swap/page across the network Put a big enough disk in to run the OS - Don’t have to mess with net booting Put a big enough disk in to run the OS and keep some data cached locally I favor the last method. Disk space is cheap, labor expensive

THE MOVES INSTITUTE CD If you do go the disk-per-compute node route you should have at least a CDROM drive in order to boot from CD and install. OS bloat has made it likely that a DVD may be required in a few years.

THE MOVES INSTITUTE Video You should have at least an el-cheapo VGA card in order to hook up a monitor. If you’re doing visualization work a high- end graphics card may be needed. Note that a fancy graphics card may consume more power and generate more heat.

THE MOVES INSTITUTE Networking The compute nodes have to communicate with each other and the front end over the private network. What network should be used for this? We want: High speed Low latency Cheap Major technologies are gigabit ethernet, Myrinet, and Infiniband

THE MOVES INSTITUTE Networking Myrinet is a 2 Gb/sec, 2-3 ms latency networking standard that uses multimode fiber NIC price (as of this writing) is about $500, and a 16 port switch about $5K.

THE MOVES INSTITUTE Infiniband Infiniband is all-singing, all-dancing, 10 Gb/Sec (for 4X rate), ~4 ms latency Pricing seems to be similar to that of Myrinet. HP prices on the order of $1K for a PCI adapter, $10K for a 24 port switch. Other places are probably cheaper. Often a cluster will have both an infiniband and an ethernet network, the Infiniband for MPI and the ethernet for conventional communications

THE MOVES INSTITUTE Gigabit Ethernet 1 Gb/S, high latency (it goes through the TCP/IP protocol stack by default) Advantage: dirt cheap, ubiquitous. Gbit ethernet is built into most server mobos by default, unmanaged L2 gigabit switches are at aprox $10/port You can reuse existing expertise, cables, etc. It’s pretty tough to argue against gigabit ethernet on an economics basis

THE MOVES INSTITUTE Front End The front end is the cluster’s face to the rest of the world. It alone is connected to the public network. It typically runs a web server, scheduling software, NFS, DHCP on the private network, a firewall, and a few other utilities More memory than a compute node is good. More disk is good. (disk is the subject of another talk.)

THE MOVES INSTITUTE Price Very roughly: Compute nodes at $3K/each (dual opterons, 4 gig mem, 120 gig disk) Front end at $4K Rack at $2K Small L2 GigE switch ($150) For aprox $40K you can get a ~10 node cluster with ~20 CPUs that has a peak performance of around GFLOPS. (this pretty much ignores disk space)