GRID MetaComputing Distributed Computing GRID MetaComputing Grid MetaComputing Distributed Computing Communications CS 594 Spring 2003 For two or more MetaComputing entities compute nodes MPPs data real time instruments printers scanners to interact they need 3 basic items Communications Naming of entities Interfaces Protocols or just standard methods for now Note than in Message Passing systems we only needed 2 out of the 3 In MetaComputing the extra one is due to the Heterogeneous nature of the system involved Thanks to Mark Baker for the use of this material Communications Communications Communication Requirements What s a computer network A set of nodes connected by communication links A set of networks connected by nodes Nodes can be general purpose computers Workstations Servers Nodes can also be more specialised systems Routers Switches Communications Connectivity Nodes communicating via links Recursive connectivity Networks communicating via gateways Addressing Identifying a node Routing Forwarding messages to a node Broadcast and Multicast Sending messages to multiple destinations at once Multiplexing Multiple parties sharing a link Reliability Recovering from errors Performance Low latency high bandwidth Communications Networks can have different topologies No matter what the physical topology the system using the network will want to be able to define its own view of the network Logical vs Physical bus ring Network Layer Environment SW view star I e MPI topology calls Ethernet broadcast vs the internet MBONE tree mesh 1 Communications Issues in networking low level Error Control recovery Multiplexing and de multiplexing joining multiple data streams or splitting one stream into multiple Fragmentation divide data into network size pieces Addressing identifying network entities Routing finding a path through the network Flow control avoid overloading slower entity Congestion control avoid overloading network Communications A complex problem can be simplified by layering Layer N relies on services provided by layer N 1 Layer N provides service to layer N 1 Interfaces between layers completely define services an API Layers hide complexity Service provided by layer is independent of layer s implementation Changes inside layer do not affect other layers Communications Protocols define the interface and interactions between entities Applications don t worry about protocols just the immediate interface to the next layer down What is a networking protocol A set of rules for how network elements communicate Specifies the format of messages exchanged Specifies actions to take on receipt of message Protocol specifications must be exact Communication Layers Application layer Provides process to process communication All other layers exist to support this layer Examples Email FTP Telnet WWW Presentation layer Converts data to a common format Little endian v big endian byte orders Size of data structures Communications Application Application 7 Presentation Presentation 6 Session Session 5 Transport Transport 4 Network Network Network 3 Data Link Data Link Data Link 2 Physical Physical Physical 1 A B C Communication Layers Session layer Binds two transport streams into a relationship e g audio video Performs authentication during session setup Transport layer End to end communication Reliable in order delivery Multiplexes higher level streams Flow and congestion control 2 Communication Layers Network layer Accepts packets from upper layers and injects them into network Specifies host addresses Implements packet routing Controls congestion discards packets if needed Communication Layers Data link layer Provides point to point error free communication over a single link Matches speed between sender and receiver Framing and error control Media Access Control MAC Physical layer Communicates raw bits Deals with voltage levels frequencies Usually the realm of electrical engineers Protocol Data Units PDUs PDUs Send side Layer N takes PDU from layer N 1 Adds its own fields header to form a new PDU Passes PDU to layer N 1 Receive side Layer N takes PDU from layer N 1 Strips and processes its own fields Passes PDU to layer N 1 Layering harmful A powerful technique for structuring systems but religious adherence has its problems Different layers may duplicate functionality Multiplexing at transport network and data link Fragmentation at transport and network Error recovery at transport and data link Congestion control at transport and data link Layering harmful Different layers may need access to same information maximum transmission unit timestamps Must compromise between modularity and performance Compare PVM internals MPI CH Nexus MPICH g MPICH g2 3 Layering How to avoid the overheads Direct access to the network hardware Memory mapped hardware Elan chip set Meiko Paragon and NX vs NX2 Layering How do we handle multiple interfaces to multiple hardware networks on the same machine Which layer does the hard work User API environment OS Naming interfaces services SAPs messages or just end points who handles the errors What about really fast interfaces eui fm VIA VBE IBM SP2 css0 vs en0 swXnY vs fXnY Application Layer Protocols A number of standard application layer protocols File Transfer Protocol FTP Simple Mail Transfer Protocol SMTP Network News Transfer Protocol NNTP HyperText Transfer Protocol HTTP You can also roll your own If you want a good meta computing system then yes you need too see PACX Transport layers Transport layers Transmission Control Protocol TCP Provides reliable byte stream service Header fields sequence numbers port numbers checksum User Datagram Protocol UDP Provides unreliable unordered datagram service Header fields port numbers checksum Typical Application Comm Layers Internet Protocol IP The key piece of the architecture Can use many different link layers Treats each network in Internetwork as a link No quality of service QoS guarantees Can lose packets and misorder packets Header fields host addresses hop counts Application TCP UDP IP Network Note no port number 4 Flow and Congestion Control Flow and Congestion Control The TCP backoff algorithm which employs rapid rate reduction on packet loss and slow rate increase in the absence of packet loss is very effective for data transmission as in the data intensive distributed supercomputing applications It allows many TCP connections to share the network fairly an efficiently without explicit co ordination TCP
View Full Document
Unlocking...