ίκτυα Pedro Trancoso Τµήµα Πληροφορικής, Πανεπιστήµιο Κύπρου Εισαγωγή Why look at networks? Shrinking networks => connection between computers, I/O devices, boards, chips, and even modules inside chips All computers are networked to other devices Clusters? Interconnected computers inside room 1
Είδη ικτύων ίκτυο Ευρείας Περιοχής - Wide area network (WAN) Thousands of computers, thousands of kilometers Τοπικό ίκτυο - Local area network (LAN) Hundreds of computers, few kilometers ίκτυο Αποθήκευσης ή Συστήµατος -Storage or system area network (SAN) Hundreds of nodes, 100 meters Απλό ίκτυο Μηνύµατα: Header+Payload+Trailer Send a Message (1) Application copies data into OS buffer (2) OS calculates checksum, includes it in header or trailer and sets timer (3) OS sends data to network interface Receive a message (3), (2), (1)! Sender reaction If gets acknowledgment, releases copy of message If gets timeout, resends data 2
Ορισµοί Bandwidth Time of flight: time for first bit to arrive Transmission time: time for message to pass through system excluding time of flight Transport latency: time of flight + transmission time Sender overhead: time to inject msg into net Receiver overhead: time to pull msg from net Ορισµοί 3
Μέγεθος των Μηνυµάτων NFS traffic Καλώδια Συνεστραµµένο Ζεύγος Καλωδίων -Twisted Pair Οµοαξονικό Καλώδιο -Coaxial Τεχνολογία Οπτικών Ινών - Fiber Optics 4
Καλώδια Συνεστραµµένο Ζεύγος Καλωδίων -Twisted Pair 10Mb/s Οµοαξονικό Καλώδιο -Coaxial Cable 10Mb/s over kilometer Τεχνολογία Οπτικών Ινών - Fiber optics Multimode fiber: inexpensive LED, 62.5-micron, 100m @ 1000Mbps or few Km @ 100Mbps Single-mode fiber: 8/9-micron, 100s Km @ Gbps Σύνδεση µε περισσότερους από δυο υπολογιστές Shared versus Switched Media Carrier sensing and collision detection Point-to-point communication Broadcast, multicast Data switching exchanges, multistage interconnection networks, interface message processors 5
Σύνδεση µε περισσότερους από δυο υπολογιστές Connection-Oriented versus Connectionless Communication Frequency-division multiplexing ροµολόγηση - Routing: Delivering Messages Source-based Routing Virtual Circuit Destination-based Routing Πολιτική - Policy Store-and-forward Cut-through or wormhole routing Σύνδεση µε περισσότερους από δυο υπολογιστές Έλεγχος Συµφόρησης -Congestion Control Packet discarding Έλεγχο Ροής - Flow control Backpressure feedback Credit-based flow control Choke packets 6
Τοπολογία ικτύου Συγκεντρωτικό Μεταγώγιµο - Centralized Switch ίκτυο ιασταύρωσης - Crossbar Omega Network Fat tree Κατανεµηµένο Μεταγώγιµο - Distributed Switch ίκτυο ακτυλίου -Ring Torus Hypercube Τοπολογία ικτύου Παράδειγµα: Consider nodes as elements of a 2-D array with nearest neighbor communication (NEWS North, East, West, South). 8-by-8 array. How long does it take to send one message to northern neighbor and one to eastern neighbor? For Bus, Fully Connected, Ring, 2D Torus, 6-Cube Messages: (64-15)*2+14*1=112 Bus: 112 Fully: 1 Ring: 9 2D Taurus: 2 7
Πρακτικά Ζητήµατα για Εµπορικά ίκτυα - Practical Issues for Commercial Interconnection Networks Συνεκτικότητα - Connectivity Connecting the Network to the Computer Standardization: Cross-Company Interoperability Message Failure Tolerance Node Failure Tolerance Ethernet: Το Τοπικό ίκτυο Γέφυρα ικτύων (Bridges); ροµολογητής ικτύων (Routers) ήπύλη ικτύων (Gateways) 8
Infiniband: ίκτυο Αποθήκευσης Clock 2.5GHz, Peak speed 2000Mbps Packet switched, connectionless network Cut-through routing, destination-based addressing Backpressure storage control Very light protocol! (e.g. no security) ATM: ίκτυο Ευρείας Περιοχής 155Mbps, 620Mbps, 2480Mbps, Fiber optics Virtual channel Store-and-forward routing Small fixed-sized packets (48-byte payload) -> quality-of-service Simple switches but large and complex protocol stack 9
ιασύνδεση ικτύων - Internetworking TCP/IP Open Systems Interconnect (OSI) Ζητήµατα ικτύων - Crosscutting Issues for Interconnection Networks Density-Optimized Processors versus SPEC- Optimized Processors Smart Switches versus Smart Interface Cards Protection and User Access to the Network Efficient Interface to Memory Hierarchy versus Interconnection Network Compute-Optimized Processors versus Receiver Overhead 10
Συµπλέγµατα -Clusters Performance Challenges Dependability and Scalability Advantage Pros and Cons of Cost Shooting for the Best of Both Worlds Popularity Amazon, AOL, Google, Hotmail, Inktomi, WebTV, and Yahoo Σχεδίαση ενός Συµπλέγµατος - Designing a Cluster 32 CPU Cluster, Cost of ownership for 3-year 11
Το Σύµπλεγµα Υπολογιστών Number of pages on the web grew by a factor of 1000 between 1994 and 1997 People want top 10 answers only! In December 2000, Google indexed 1327 million pages, had 70 million queries-per-day Requirements: Reliability: continuously available Small latency: 0.5s High bandwidth: 1000 queries per second Crawling the Web Το Σύµπλεγµα Υπολογιστών December 2000: More than 6000 processors, 12.000 disks (1 petabyte storage) No RAID, redundant sites! Performance Bandwidth: Query 26Mbps, Crawl 53Mbps, Replication 132Mbps! Cost Best cost-performance CPU not latest model! Large energy demands Reliability Software: 20 reboot per day Hardware: 2-3% PC replacement per year (disk, DRAM failures) 12
Το Σύµπλεγµα Υπολογιστών Μέσα από ένα Κινητό Τηλέφωνο Wireless networks Cells Cell phone = mobile computer acting as a radio! I/O devices: keyboard and LCD display, etc. Protocol to find cell 13
Μέσα από ένα Κινητό Τηλέφωνο Standards: CDMA AMPS TDMA GSM Πλάνες και Παγίδες (Fallacies and Pitfalls) Pitfall: Using bandwidth as the only measure of network performance Pitfall: Ignoring software overhead when determining performance Intel Paragon: HW 0.2us SW 250us! 25us Pitfall: Trying to provide features only within the network versus end to end Pitfall: Relying on TCP/IP for all networks, regardless of latency, bandwidth, or software requirements 14
Συµπεράσµατα Internet and WWW revolutionize information access: TCP/IP makes it practical Switches is a growing technology Network Processors Networking at all levels: within chip, across chips, across boards, etc. Disks connected with a SAN Clusters of small computers connected with LAN Performance grows 15