Distributed Systems. Ordering of events. Atomicity. State. Competition for resources. Selecting a co-ordinator. Architectures. Distributed clocks

Transcription

1 By 'geographic region' IP Addresses Giving things unique identifiers Handed out by a server? URLs DNS By 'vendor' MAC addresses stateless systems State stateful systems The Web Ditributed Games Peer to peer Architectures Client / server Thick clients Thin clients Bully algorithm Selecting a co-ordinator Distributed Systems Two Phase Commit Selecting a master node for peer-to-peer e.g. p2p games Deadlock Livelock Detecting it when it happens? Avoiding it in the first place? Competition for resources Starvation Dining philosophers Implies resources can be 'locked' Mutual exclusion ('mutex') Atomicity Critical sections Two-phase committ Lamport Clocks Ordering of events Distributed clocks Berkley algorithm Christian's algorithm

2 Centralised? Not really feasible Assigning unique identifiers MAC Addresses IP Addresses 'The Internet' DNS Names Random? Hierarchical Could sort-of work if the name space is big enough? By geographic region By 'vendor 'Infrastructure' (e.g. DNS) URLs Different transport protocols Stateless system Resources identified by URLS TCI/IP for 'slow' / 'reliable' comms UDP for 'fast' / 'unreliable' comms Static resources such as files and directories Dynamic resources (blogs / oneline-stores / games) The Web Thin or thick? Thin : because application logic is mostly on the server Thick : AJAX / Flash etc run in the browser Can only achieve state 'per-application' Cookies POST / GET Inherently 'stateful' ( s are themselves packets of state) Distributed Systems Sending mail Receiving mail SMTP Doesn't have authentication POP3 IMAP Clients have a local cache of s Try to achieve a kind of 'distributed causality' Games Not all state is important Multiple players see the same 'world' Has to be consistent or the game makes no sense Heavily stateful Prioritise syncing of state that is critical to game logic Allow non-critical state to become out-of-sync if resources are limited

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

4 MN,A$%%)$&&)#'.1%)(1..13&-*%'4.14'61+MN/44+'%%$37?%%$73'4/))#')1:&','&0-)#'M3)'+3')?%%$73'4T*.0'+%?*)#1+$)-?34)#'34'&'7/)'4)1+'7$13%?34)#'34'&'7/)'4)1(1*3)+$'%2$3%)$)*)$13%')(5 ;.<=3/)>=<06)

5 7)2CD31=E)FD93B=20G)/02H.9I)602AD)

6 Universal Resource Identifier (URI) Universal Resource Name (URN) Universal Resource Locator (URL)

7 Mail Server n EightbasicHTTPmethods GET3fetcharesource srepresentation(whichincludesa header ) HEAD3getjustaresource sheader POST3submitsomedatatobeprocessed PUT3uploadsaresource DELETE3deletesaresource TRACE3echoesarequestbacktotheclient,fordebuggingpurposes OPTIONS3returnsthemethodsthataparticularserversupports CONNECT3complicatedconnectivitystuff ServersmustimplementGET,HEADandOPTIONS

8 Example POP3 session S: <wait for connection on TCP port 110> C: <open connection> S: +OK POP3 server ready C: APOP mrose c4c9334bac560ecc979e58001b3e22fb S: +OK mrose's maildrop has 2 messages (320 octets) C: STAT S: +OK C: LIST S: +OK 2 messages (320 octets) S: S: S:. C: RETR 1 S: +OK 120 octets S: <the POP3 server sends message 1> S:. C: DELE 1 S: +OK message 1 deleted C: RETR 2 S: +OK 200 octets S: <the POP3 server sends message 2> C: QUIT S: +OK dewey POP3 server signing off (maildrop empty) C: <close connection> S: <wait for next connection>

9 Dining Philosophers Think for a while Get hungry Get BOTH chopsticks Eat for a while Put chopsticks back where they came from IMAP Internet Messaging Access Protocol Connected and Disconnected modes - cache like Multiple clients connected to same mailbox Multiple mailboxes Server side searches

10 telnet capone.rutgers.edu 143 Trying Connected to Escape character is '^]'. * OK [CAPABILITY IMAP4REV1 LOGIN-REFERRALS STARTTLS AUTH=LOGIN] capone.rutgers.edu IMAP4rev at Wed, 13 Apr :38: (EDT) A1 LOGIN mailtest Password A1 OK [CAPABILITY IMAP4REV1 IDLE NAMESPACE MAILBOX-REFERRALS BINARY UNSELECT SCAN SORT THREAD=REFERENCES THREAD=ORDEREDSUBJECT MULTIAPPEND] User mailtest authenticated A2 SELECT Inbox * 2 EXISTS * 2 RECENT * OK [UIDVALIDITY ] UID validity status * OK [UIDNEXT 3] Predicted next UID * FLAGS (\Answered \Flagged \Deleted \Draft \Seen) * OK [PERMANENTFLAGS (\* \Answered \Flagged \Deleted \Draft \Seen)] Permanent flags * OK [UNSEEN 1] first unseen message in /var/mail/mailtest A2 OK [READ-WRITE] SELECT completed * BYE capone.rutgers.edu IMAP4rev1 server terminating connection A3 FETCH 2 BODY[HEADER] * 2 FETCH (BODY[HEADER] {670} Return-Path: X-Original-To: mailtest@capone.rutgers.edu Delivered-To: mailtest@capone.rutgers.edu Received: from node18.rutgers.edu (node18 [ ]) by capone.rutgers.edu (Postfix) with ESMTP id A291B2B15C for ; Tue, 12 Apr :23: (EDT) Received: from me?here.com (unknown [ ]) by node18.rutgers.edu (Postfix) with SMTP id 4653B14112 for ; Tue, 12 Apr :24: (EDT) To: some_guru@somewhere.com From: pp@pp.com Subject: Forged Message-Id: < B14112@node18.rutgers.edu> Date: Tue, 12 Apr :24: (EDT) ) * 2 FETCH (FLAGS (\Recent \Seen)) A3 OK FETCH completed

11 A4 FETCH 2 BODY[TEXT] * 2 FETCH (BODY[TEXT] {88} Hey, The "To:" and "From:" are non-existent, but you still get the . bye, bye ) A4 OK FETCH completed A5 LOGOUT * BYE capone.rutgers.edu IMAP4rev1 server terminating connection A5 OK LOGOUT completed Connection closed by foreign host. SMTP Simple Mail Transfer Protocol Text based Takes a mail message and some header information and sends it to a mail server

12 S: 220 smtp.example.com ESMTP Postfix C: HELO relay.example.org S: 250 Hello relay.example.org, I am glad to meet you C: MAIL FROM:<bob@example.org> S: 250 Ok C: RCPT TO:<alice@example.com> S: 250 Ok C: RCPT TO:<theboss@example.com> S: 250 Ok C: DATA S: 354 End data with <CR><LF>.<CR><LF> C: From: "Bob Example" <bob@example.org> C: To: Alice Example <alice@example.com> C: Cc: theboss@example.com C: Date: Tue, 15 Jan :02: C: Subject: Test message C: C: Hello Alice. C: This is a test message with 5 headers and 4 lines in the body. C: Your friend, C: Bob C:. S: 250 Ok: queued as C: QUIT S: 221 Bye {The server closes the connection}

13 p5 q3 p4 p3 q2 p2 p1 q1 If A precedes B, then we write A B If A and B are concurrent events, we can t say anything about their overall ordering But we can use message passing to establish at least a partial happened before relationship

14 Logical Clocks If each processor has a logical clock LCi When an event occurs, increment LCi When processor X sends a message to Y, it also sends LCx When Y receives the message if LCy < (LCx +1): LCy = LCx + 1 if a b then it is true that LCa < LCb BUT With a Lamport Clock... if LCa < LCb it is not necessarily true that a b This means that we can t infer a causal ordering just by looking at the timestamps

15 Co-ordinating Generals Take any exchange of messages that appears to end up in a co-ordinated solution Assume the last message gets lost; if the scenario still succeeds, then this last message wasn t actually vital to the scenario Repeat until no messages are exchanged No messages clearly isn t a solution Our initial assumption about the solution is therefore clearly wrong

16 Mutual Exclusion ( mutex ) A centralised version Client: Send a request to the server for a lock on the mutex When reply comes back, start the critical section When finished, send a message to release the mutex The Lock Server If the mutex is available, mark it as being used by that client, and send a reply. Otherwise, queue the request When the mutex is released by the client, if someone else wants it, pass it to them, otherwise mark it as being available

17 A distributed version A process K wants to enter a critical section: Generates a timestamp TS k Sends a request (K, TS k) to everyone Waits for a reply from everyone that it s ok to proceed Enters its critical section When a process receives a request... If it s already in its critical section, it defers its reply Otherwise, if it doesn t itself want to enter the critical section, it sends a reply If we want to enter the critical section, if the requestor has a smaller timestamp than us, allow them in first, otherwise defer the reply A distributed version A process K wants to enter a critical section: Generates a timestamp TS k Sends a request (K, TS k) to everyone Waits for a reply from everyone that it s ok to proceed Enters its critical section When a process receives a request... If it s already in its critical section, it defers its reply Otherwise, if it doesn t itself want to enter the critical section, it sends a reply If we want to enter the critical section, if the requestor has a smaller timestamp than us, allow them in first, otherwise defer the reply

18 Can we arrive at a state where we can guarantee that either something has happened correctly, or has had definitely NOT happened? Two Phase Commit

19 2PC : Phase 1 A co-ordinator requests a transaction, and sends a request to all participants e.g. to C1 sends request to remove X pounds from account, and to C2 sends request to add X pounds to account All participants respond as to whether they are willing and able to execute the request, and send VOTE_COMMIT or VOTE_ABORT They log their current state, and then perform the transaction All participants log their vote 2PC : Phase 2 The co-ordinator looks at the votes If everyone has voted to commit, then the coordinator sends a GLOBAL_COMMIT to everyone; otherwise it sends a GLOBAL_ABORT On receiving the decision from the co-ordinator, all participants record the decision locally If it was an ABORT, participants ROLL BACK to their previous safe state

20 Which node is the coordinator? If we had a central server we probably wouldn t have to worry about a lot of these problems... but we d have other problems. So, in a distributed system, how can we arrange for one node to take charge of algorithms such as 2PC? Bully Algorithm (Garcia-Molina, 1982) is a mechanism for choosing a co-ordinator from a set of candidate nodes Relies on some ordering / numbering of nodes 1. P sends an ELECTION message to all processes with higher numbers 2. If no one responds, P wins the election and becomes the co-ordinator 3. If one of the higher-numbered nodes answers, it takes over

21 Election Election 6 Election 0 3 7

22 I m still here I m still here Election 4 0 Election Election 3 6 7

23 1 2 5 I m still here I m the new co-ordinator

24 Limitations? A time-out is required (i.e. If no one responds.. but for how long?) Requires an initial ordering of nodes; this isn t always easy to achieve (why?) Game Logic Fire sheep with specific velocity Move in a straight line until you hit the trampoline Invert vertical direction Check to see if sheep lands in bucket

25 Sheep Trajectory Version 1 Hold x and y co-ordinates for sheep Decide on dx and dy At each timestep x = x + dx y = y + dy Negate dy when trampoline has been hit Sheep Trajectory Version 2 Hold x and y co-ordinates for sheep Decide on dx and dy At each timestep x = x + dx * t y = y + dy * t now what happens when we hit the trampoline?

26 Cristian s Algorithm P requests the time from S S prepares a response that the time is T, and sends it at the last possible moment P sets its time to be T + RoundTripTime / 2 Problems? What happens if the round trip time is different when the update is sent to that which was measured? What happens if send and receive don t take the same amount of time?

27 Berkeley algorithm Choose a master node Master then polls the slaves, who reply with their time observes the RTT and estimates the time in each slave and its own time averages the clock times, ignoring any outlying values sends each slave a relative update to their clock, which can be +ive or -ive Problems? What happens to a node that s had its clock set backwards?