1. EEC-681/781 Distributed Computing Systems Lecture 4 Wenbing Zhao Department of Electrical and Computer Engineering Cleveland State University wenbing@ieee.org

2. 2 Fall Semester 2006EEC-681: Distributed Computing SystemsWenbing Zhao Outline Architecture Model Server load balancing End-to-end arguments in system design Inter-process communications (part 1)

3. 3 Fall Semester 2006EEC-681: Distributed Computing SystemsWenbing Zhao Architectural Models Client-server model Peer-to-peer model

4. 4 Fall Semester 2006EEC-681: Distributed Computing SystemsWenbing Zhao Basic Client–Server Model

5. 5 Fall Semester 2006EEC-681: Distributed Computing SystemsWenbing Zhao Servers Servers: Generally provide services related to a shared resource: –Servers for file systems, databases, implementation repositories, etc. –Servers for shared, linked documents –Servers for shared applications –Servers for shared distributed objects

6. 6 Fall Semester 2006EEC-681: Distributed Computing SystemsWenbing Zhao Client Clients: Allow remote service access –Programming interface transforming client’s local service calls to request/reply messages –Devices with (relatively simple) digital components (barcode readers, teller machines, hand-held phones) –Computers providing independent user interfaces for specific services –Computers providing an integrated user interface for related services (compound documents)

7. 7 Fall Semester 2006EEC-681: Distributed Computing SystemsWenbing Zhao Application Layering Traditional three-layered view: –User-interface layer contains units for an application’s user interface –Processing layer contains the functions of an application, i.e. without specific data –Data layer contains the data that a client wants to manipulate through the application components

8. 8 Fall Semester 2006EEC-681: Distributed Computing SystemsWenbing Zhao Multitiered Architecture

9. 9 Fall Semester 2006EEC-681: Distributed Computing SystemsWenbing Zhao Multitiered Architecture 1-30

10. 10 Fall Semester 2006EEC-681: Distributed Computing SystemsWenbing Zhao Horizontal Distribution 1-31

11. 11 Fall Semester 2006EEC-681: Distributed Computing SystemsWenbing Zhao Modern Architectures Vertical distribution –According to application logic Horizontal distribution –Scalability and availability In practice, a system is distributed in both directions

12. 12 Fall Semester 2006EEC-681: Distributed Computing SystemsWenbing Zhao Peer-to-Peer Model A different way to construct client-server systems where most, or all, of the server functionality resides on the clients themselves Advantages –Scalable to very large numbers (millions) –Stable under very high load –Self-repairing when disruptive failures occur

13. 13 Fall Semester 2006EEC-681: Distributed Computing SystemsWenbing Zhao Why Server Load Balancing? Scalable - Upgrade out instead of up Easily grow - Spread your load across your existing low-cost servers as your needs grow HA - Can work closely with High Availability solutions to add and remove servers as necessary Slides on server load balancing taken from Dustin Puryear’s talk on “LVS – Load Balancing and High Availability for Free” at USENIX LISA 2003

14. 14 Fall Semester 2006EEC-681: Distributed Computing SystemsWenbing Zhao SLB Layout

15. 15 Fall Semester 2006EEC-681: Distributed Computing SystemsWenbing Zhao SLB Definitions Virtual IP (VIP) – The IP used by clients to access a service Load Balancer (LB) or Server Load Balancer (SLB) – the server that balances packets going to or from the servers providing the service Real Server – The server providing the actual service (i.e., the server running Apache) Real IP (RIP) – The IP of each real server Schedule – How the SLB determines which real server gets the next connection

16. 16 Fall Semester 2006EEC-681: Distributed Computing SystemsWenbing Zhao SLB Transaction Walkthrough 1.Client established TCP connection with SLB. 2.SLB determines which real server is ready for a connection using configured schedule, and creates a connection to that server. 3.Client sends HTTP request. 4.SLB forwards HTTP request to real server. 5.Real server responds back to.. Who?

17. 17 Fall Semester 2006EEC-681: Distributed Computing SystemsWenbing Zhao So How Exactly Does the Real Server Respond? There are two traditional methods to solving this problem: –Network Address Translation (NAT) –Direct Routing/Direct Server Return

18. 18 Fall Semester 2006EEC-681: Distributed Computing SystemsWenbing Zhao NAT-Based SLB

19. 19 Fall Semester 2006EEC-681: Distributed Computing SystemsWenbing Zhao NAT-Based SLB SLB sits between the client and real servers for all incoming and outgoing packets, just like a router that does NAT The SLB can rewrite the IP packet in one of two ways: –Half-NAT: load-balancer rewrites the destination IP –Full-NAT: load-balancer rewrites the source and destination IP The VIP is only assigned on the load-balancer

20. 20 Fall Semester 2006EEC-681: Distributed Computing SystemsWenbing Zhao Half-NAT 1.SLB gets packet from client 2.SLB changes the destination IP in the header of the packet 3.SLB forwards packet to real server Major benefit of this method is that the real server knows the real IP of the client –The real server can properly log client IP’s –The real server can do better security based on the IP address of the client

21. 21 Fall Semester 2006EEC-681: Distributed Computing SystemsWenbing Zhao Full-NAT 1.SLB gets packet from client. 2.SLB changes the destination IP and source IP in the header of the packet. 3.SLB forwards packet to real server. Major drawback to this technique: –The real server has no idea who sent the original request

22. 22 Fall Semester 2006EEC-681: Distributed Computing SystemsWenbing Zhao Direct Routing/Direct Server Return Layer 2 LB

23. 23 Fall Semester 2006EEC-681: Distributed Computing SystemsWenbing Zhao Direct Routing/Direct Server Return 1.SLB gets packet from client 2.SLB changes destination IP in packet header 3.SLB forwards packet to real server, which also has VIP 4.Real server responds directly to the client with VIP as the source IP There are benefits and costs to this approach –Without having to push data back through the SLB you can decrease latency –How do you handle ARP if multiple devices have the same IP (the VIP)?

24. 24 Fall Semester 2006EEC-681: Distributed Computing SystemsWenbing Zhao End-to-End Arguments in System Design Concerned with placement of functions among modules of a distributed system Layered systems are very common –Network protocols and Middleware Claim –Functions at low levels of a system may be redundant or of little use –Moving a function upward in a layered system closer to application that uses the function

25. 25 Fall Semester 2006EEC-681: Distributed Computing SystemsWenbing Zhao Careful File Transfer Objective: move a file from computer A’s storage to computer B’s storage Application program: file transfer program

26. 26 Fall Semester 2006EEC-681: Distributed Computing SystemsWenbing Zhao Careful File Transfer Fault-free steps of file transfer –A reads the file from its disk –A sends the file to its networking stack and the file is packetized at the networking protocol stack –Communication network moves packets from A to B –At B, B’s networking protocol stack delivers transmitted file to the file transfer program running in B –B’s file transfer program asks its file system to write the file to B’s disk

27. 27 Fall Semester 2006EEC-681: Distributed Computing SystemsWenbing Zhao Careful File Transfer What can go wrong in each step –File corrupted due to disk error –Software bugs in buffering and copying the file, in A or B –Processor or memory transient error during buffering or copying –Communication system error, e.g., drop or change bits, or duplicate delivery –Host might crash

28. 28 Fall Semester 2006EEC-681: Distributed Computing SystemsWenbing Zhao Careful File Transfer How to cope with the threats –Reinforce each step along the way using time-out and retry, error detection, crash recovery, etc. –Doing everything three times

29. 29 Fall Semester 2006EEC-681: Distributed Computing SystemsWenbing Zhao End-to-end Check and Retry Assuming low probability of the threats As a final additional step –B reads back the file from disk into memory –Recalculates checksum –Sends this value back to A –A compares the value with checksum of original file –If two checksums agree, file transfer completes. Otherwise, retry from beginning

30. 30 Fall Semester 2006EEC-681: Distributed Computing SystemsWenbing Zhao Is Reliable Communication Useful? Communication system can be made reliable with Packet checksums, sequence number, internal retry, etc. Is it enough to ensure correct file transfer? –No. Other threats can still corrupt the file –The extra end-to-end check and retry still necessary –Usefulness: Reducing frequency of retry

31. 31 Fall Semester 2006EEC-681: Distributed Computing SystemsWenbing Zhao Is Reliable Communication Useful? Conclusion –Communication system to go out of its way to be extraordinarily reliable does not reduce the burden on the application program to ensure reliability

32. 32 Fall Semester 2006EEC-681: Distributed Computing SystemsWenbing Zhao Performance Aspects Cannot conclude that lower levels should play no part in obtaining reliability –If file is long, retry the transfer of the file is too expensive –By providing reliable communication, can significantly enhance the performance

33. 33 Fall Semester 2006EEC-681: Distributed Computing SystemsWenbing Zhao Performance Aspects Putting reliability measures in lower level –An engineering trade-off based on performance –Not a requirement for correctness –Must be careful. Might not always improve performance Other application might not need the extra functionality Might not have enough information to operate efficiently

34. 34 Fall Semester 2006EEC-681: Distributed Computing SystemsWenbing Zhao Inter-Process Communications Techniques: –Shared memory –Message passing Objectives: –Data exchange –Synchronization: processes at different hosts, executing at different rates, need to influence the overall execution pattern => Constraints on the order of events

35. 35 Fall Semester 2006EEC-681: Distributed Computing SystemsWenbing Zhao The OSI Network Architecture

36. 36 Fall Semester 2006EEC-681: Distributed Computing SystemsWenbing Zhao Low-level Layers Physical layer: contains the specification and implementation of bits, and their transmission between sender and receiver Data link layer: prescribes the transmission of a series of bits into a frame to allow for error and flow control Network layer: describes how packets in a network of computers are to be routed

37. 37 Fall Semester 2006EEC-681: Distributed Computing SystemsWenbing Zhao Transport Layer The transport layer provides the actual communication facilities for most distributed systems. TCP: connection-oriented, reliable, stream-oriented communication UDP: unreliable (best-effort) datagram communication

38. 38 Fall Semester 2006EEC-681: Distributed Computing SystemsWenbing Zhao Application Layer Many application protocols are directly implemented on top of transport protocols that do a lot of application- independent work FTPWWW TransferFTPHTTP Encoding7-bit text + 8-bit binary8-bit + content type NamingHost + pathURL DistributionPull ReplicationCaching + DNS tricks

39. 39 Fall Semester 2006EEC-681: Distributed Computing SystemsWenbing Zhao Message Layout

40. 40 Fall Semester 2006EEC-681: Distributed Computing SystemsWenbing Zhao Middleware Layer Middleware is invented to provide common services and protocols that can be used by many different applications: –A rich set of communication protocols –Marshaling and unmarshaling of data –Naming protocols –Security protocols –Scaling mechanisms What remains are truly application-specific protocols

41. 41 Fall Semester 2006EEC-681: Distributed Computing SystemsWenbing Zhao An Adapted Reference Model with Middleware Layer 2-5

42. 42 Fall Semester 2006EEC-681: Distributed Computing SystemsWenbing Zhao Duplicate Message Suppression At networking stack –A message may be sent twice as a result of time-out-triggered failure detection and retry mechanisms –If duplicate message is suppressed here, does it mean that the application no longer needs to watch for duplicates? No

43. 43 Fall Semester 2006EEC-681: Distributed Computing SystemsWenbing Zhao Duplicate Message Suppression Application itself might originate duplicate message –To cope with suspected failures using a retry mechanism –These messages appear to be different messages to the networking stack End-to-end arguments –Application still has to check and suppress duplicates –Is low-level duplicate suppression a must if the message is large?