Documente Academic
Documente Profesional
Documente Cultură
calculatoarelor
Curs Calculatoare/Electronica
aplicata anul III
2013
Ce reprezinta arhitectura si
organizarea calculatoarelor?
• Arhitectura exprima acele Nivele de
caracteristici vizibile structurare
In cadrul sistemelor
programatorului, principalul de calcul:
•Sisteme de operare
element il reprezinta ISA •Programe monitor
•Programe dedicate
(Instruction Set Architecture) – •Nivel fizic
• Fluxul de intrare
• Fluxul de
procesare/memorare
• Fluxul de iesire
Detalierea fluxurilor de procesare
Posibilitatile de transfer a
informatiilor si procesare (1/2)
• Transferul informatiilor:
– Zone de memorie comuna (MC):
Acest tip de transfer presupune accesarea succesiva a aceleiasi zone de memorie
dinspre cel putin doua magistrale avand capacitati de transfer diferite. Rolul acestui
tip de transfer MC consta in asigurarea sincronizarii transferului datelor si reducerea
latentelor ce apar in procesele de transfer. In acest caz sincronizarea accesului la
date este “explicita”, iar natura datelor nu este in mod necesar cunoscuta
– Distribuita:
Corespunde sistemelor multiprocesor intreconectate in retea topologia retelelor
poate fi stelata, de tip “mesh” sau mixta. In acest caz, fiecare CPU al retelei
dispune de propriul sistem de operare, poate fi identificat de catre programator si
poate rula aplicatii atat ce tin de task-urile etelei, cat si independent de acestea
– Cluster-izata:
Corespunde sistemelor multiprocesor in organizate in retea dar care datorita
sistemului de operare si a facilitatilor acestuia de control, sunt vazute din punctul
de vedere al programatorului ca un singur calculator. Desi admit prin metoda
virtualizarii multiple sisteme de operare, totusi preponderent sunt utilizate
impreuna CPU-urile cluster-ului
Sincronizarea functionarii
calculatoarelor cu procesele controlate
• Prin scanarea porturilor sistemului:
– Presupune citirea ciclica continua a porturilor de intrare, compararea starii
acestora cu starile “marcate”-de interes prin rularea programului corespunzator
– Adoptand scanarea soft ocupam in intregime timpul CPU
• Prin Software
Registrele generale
Flux de date si unitatea Rezultate
Aritmetico - Logica
Modul de procesare digitala a
datelor
Coada
Acumulator
Memorie
Interfete de
Intrare / Iesire
Registru tampon
ALU de Memorie
Registru tampon
de Instructiuni Registru PC
Unitatea
Aritmetico
Intrare / Iesire
Interfetele de
Logica
Principala
Memoria
Unitatea
de control
a
programelor
Intern Extern
Single Instruction Multiple Data
memorie comuna
Single Instruction Multiple data
memorie distribuita
Multiple Instructions Multiple Data
Memorie comuna partajata
Masini distribuite procesare si
memorie
Arhitectura simetrica multiplu procesor
(NUMA-Non Uniform Memory Access)
AMD Opteron procesor dual core
87369245
Ultimul clasament plaseaza
China pe primul loc
7,168 NVIDIA® Tesla™ M2050 GPU
14,336 CPU
2.507 petaflops – Performanta in calcul
=2.507 x 1015
Memoria 229376 GB
Tianhe-1A
National Supercomputer Center in Tianjin
Road map HyperComputers (1)
Road Map Hyper Computers (2)
GPU versus CPU
Bull NovaScale
C-DAC Param
Cray XD1/XT3
2007 cele mai performante
computer-e cat si unde?
• BlueGene/L
DOE’s Lawrence Livermore National Laboratory in Livermore
• 360TFlops
http://www.top500.org/
Clasament Noiembrie 2010
Performantele Cray CX1 1 PFlops
• Chassis Enclosure Form Factor: 7U modular enclosure (Dimensions: W 12.22" (31.04cm) x H 17.5"
(44.45cm) x D 35.5" (90.42cm) Weight:
• - Chassis with all I/O modules (Gigabit and InfiniBand) and power supplies (28.3kg)
• Power Supplies 1600 watt hot-plug power supplies - Based on high efficiency and “power factor
correction” 1600W @ 220V = 6.66A (92% efficiency: 7.19A)
• - Redundant power supplies support 2+2 Cooling Fans Chassis comes standard with hot pluggable,
redundant fan modules based on Smart Energy Technologies Input Device Front control panel with
touch screen graphical LCD - Supports initial configuration wizard –
• Local server nodes, enclosure, and module information Two USB "pass-throughs" in front and back
Enclosure I/O Modules Up to two fabrics, featuring Ethernet switches providing uplink scalability
and high speed InfiniBand modular switches (8 ports SDR, 12 or 24 ports DDR) Ethernet Switch 16
RJ-45 auto-sensing 10/100/1000 Mbps UTP ports - Bandwidth: 32 Gbps (non-blocking) - Forwarding
Mode: Store-and-forward - Forward rate: 10 Mbps port = 14,800 packets/sec; 100 Mbps port =
148,000 packets/sec; 1000 Mbps port = 1,488,000 packets/sec - Latency: 100 to 100 Mbps = 40 μ
(max); 1000 to 1000 Mbps = 10 μ (max)
• IBS12DDR/IBS24DDR 12 or 24 Ports InfiniBand Switch IBS12DDR - Twelve -4X 10/20Gbps
(SDR/DDR) CX4 ports with support for optical adapters and cables - Subnet management software -
Embedded management with Linux OS with Ethernet and secure shell access - Ultra-low latency <
180ns - InfiniBand v1.2 compliant - Dual redundant auto-sensing - Status LED: Power, activity and
link indicators for each port, link and speed indicators built into each CX4 port - IBS24DDR -
Optional 12-4X 10/20Gbps CX4 Ports daughter card (for a total of 24 ports) - 480Gb/s (SDR) or
960Gb/s (DDR)
• Management Web-based Remote System Management Graphical Mode Console Redirection
Performance monitoring Remote control: Graceful power shutdown and reboot, hard power
shutdown and
Top500.org XI-2012 (1)
Top 500.org XI-2012 (2)