Today. Vector Clocks and Distributed Snapshots. Motivation: Distributed discussion board. Distributed discussion board. 1. Logical Time: Vector clocks

Transcription

1 Vector Clocks and Distributed Snapshots Today. Logical Time: Vector clocks 2. Distributed lobal Snapshots CS 48: Distributed Systems Lecture 5 Kyle Jamieson 2 Motivation: Distributed discussion board Distributed discussion board Users join specific discussion groups Each user runs a process on a different machine Messages (posts or replies) sent to all users in group K K oal: Ensure replies follow posts Non-goal: Sort posts and replies chronologically Can Lamport Clocks help here?

2 Lamport Clock-based discussion board Lamport Clock-based discussion board ost ost 2 C(2) = 2 eply 2 C(2) = 2 ost C() = 0 C() = 0 Want: Defer showing eply until ost arrives Time à Want: Don t defer! Time à Defer showing message if C(message) > local clock +? 5 No! ap could be due to other independent posts 6 Lamport Clocks and causality roblem generalizes: eplies to replies to posts intermingle with replies to posts Lamport clock timestamps don t capture causality iven two timestamps C(a) and C(z), want to know whether there s a chain of events linking them: a à b à... à y à z Chain of events captures replies to posts in our example Vector clock: Introduction ne integer can t order events in more than one process So, a Vector Clock (VC) is a vector of integers, one entry for each process in the entire distributed system Label event e with VC(e) = [c, c 2, c n ] Each entry c k is a count of events in process k that causally precede e 7 8 2

3 Vector clock: Update rules Initially, all vectors are [0, 0,, 0] Two update rules:. For each local event on process i, increment local entry c i 2. If process j receives message with vector [d, d 2,, d n ]: Set each local entry c k = max{c k, d k } Increment local entry c j Vector clock: Example All processes VCs start at [0, 0, 0] Applying local update rule Applying message rule Local vector clock piggybacks on interprocess messages a b [,0,0] [2,0,0] c d 2 e [2,,0] [2,2,0] f hysical time [0,0,] [2,2,2] 9 0 Comparing vector timestamps ule for comparing vector timestamps: V(a) = V(b) when a k = b k for all k V(a) < V(b) when a k b k for all k and V(a) V(b) Concurrency: a b if a i < b i and a j > b j, some i, j Vector clocks establish causality V(w) < V(z) then there is a chain of events linked by Happens-efore (à) between a and z If V(a) V(w) then there is no such chain of events between a and w [,0,0] w [2,0,0] x 2 a y z [0,,0] [2,,0] [2,2,0] 2

4 Two events a, z Lamport clocks: C(a) < C(z) Conclusion: None Vector clocks: V(a) < V(z) Conclusion: a à à z Vector clock timestamps tell us about causal event relationships VC application: Causally-ordered bulletin board system 0 riginal post VC 0 = (,0,0) VC 0 = (,,0) m s reply VC = (,,0) 2 VC 2 = (0,0,0) VC 2 = (,0,0) VC = (,,0) 2 User 0 posts, user replies to 0 s post; user 2 observes m* hysical time à 4 Today. Logical Time: Vector clocks Distributed Snapshots What is the state of a distributed system? 2. Distributed lobal Snapshots Chandy-Lamport algorithm easoning about C-L: Consistent Cuts San Francisco acct balance = $000 acct2 balance = $2000 New ork acct balance = $000 acct2 balance = $

5 Example of a global snapshot ut that was easy In our system of world leaders, we were able to capture their state (i.e., likeness) easily Synchronized in space Synchronized in time How would we take a global snapshot if the leaders were all at home? What if bama told Trudeau that he should really put on a shirt? This message is part of our system state! 7 8 System model N processes in the system with no process failures Each process has some state it keeps track of There are two first-in, first-out, unidirectional channels between every process pair and Call them channel(, ) and channel(, ) The channel has state, too: the set of messages inside For today, assume all messages sent on channels arrive intact and unduplicated lobal snapshot is global state Each distributed application has a number of processes (leaders) running on a number of physical servers These processes communicate with each other via channels A global snapshot captures. The local states of each process (e.g., program variables), along with 2. The state of each communication channel

6 Why do we need snapshots? Checkpointing: estart if the application fails Collecting garbage: emove objects that don t have any references Detecting deadlocks: The snapshot can examine the current application state rocess A grabs Lock, grabs 2, A waits for 2, waits for ther debugging: A little easier to work with than printf Just synchronize local clocks? Each process records state at some agreed-upon time ut system clocks skew, significantly with respect to CU process clock cycle And we wouldn t record messages between processes Do we need synchronization? What did Lamport realize about ordering events? 2 22 System model: raphical example Let s represent process state as a set of colored tokens Suppose there are two processes, and : rocess : rocess : When is inconsistency possible? Suppose we take snapshots only from a process perspective Suppose snapshots happen independently at each process channel(, ) Let s look at the implications... channel(, ) Correct global snapshot = Exactly one of each token

7 roblem: Disappearing tokens, put tokens into channels, then snapshot This snapshot misses,, and tokens roblem: Duplicated tokens snapshots, then sends receives, then snapshots This snapshot duplicates the token = { } = {, } 25 = {, } = {,,,, } 26 Idea: Marker messages What went wrong? We should have captured the state of the channels as well Let s send a marker message to track this state Distinct from other messages Channels deliver marker and other messages FIF Chandy-Lamport algorithm: verview We ll designate one node (say ) to start the snapshot Without any steps in between, :. ecords its local state ( snapshots ) 2. Sends a marker on each outbound channel Nodes remember whether they have snapshotted n receiving a marker, a non-snapshotted node performs steps () and (2) above

8 Chandy-Lamport: Sending process snapshots and sends marker, then sends Send ule: Send marker on all outgoing channels Immediately after snapshot efore sending any further messages Chandy-Lamport: eceiving process (/2) At the same time, sends orange token Then, receives marker eceive ule (if not yet snapshotted) n receiving marker on channel c record c s state as empty channel(,) = { } snap: = {, } 29 = {, } = {,, } 0 Chandy-Lamport: eceiving process (2/2) sends marker to receives orange token, then marker eceive ule (if already snapshotted): n receiving marker on c record c s state: all msgs from c since snapshot channel(,) = { } Terminating a snapshot Distributed algorithm: No one process decides when it terminates Eventually, all processes have received a marker (and recorded their own state) All processes have received a marker on all the N incoming channels (and recorded their states) Later, a central server can gather the local states to build a global snapshot = {, } channel(,) = { } = {,, } 2 8

9 Today. Logical Time: Vector clocks 2. Distributed lobal Snapshots Chandy-Lamport algorithm easoning about C-L: Consistent Cuts lobal states and cuts lobal state is a n-tuple of local states (one per process and channel) A cut is a subset of the global history that contains an initial prefix of each local state Therefore every cut is a natural global state Intuitively, a cut partitions the space time diagram along the time axis Cut = { The last event of each process, and message of each channel that is in the cut } Inconsistent versus consistent cuts A consistent cut is a cut that respects causality of events Consistent versus inconsistent cuts A C D A cut C is consistent when: For each pair of events e and f, if:. f is in the cut, and 2. e à f, then, event e is also in the cut 2 H E F Consistent: H à F and H in the cut Inconsistent: à D but only D is in the cut 5 6 9

10 C-L returns a consistent cut C-L can t return this inconsistent cut A C D A C D 2 E F C-L can t return this cut 2 E F H H Inconsistent: à D but only D is in the cut C-L ensures that if D is in the cut, then is in the cut 7 8 Friday recept: Cs in o Monday Topic: Eventual Consistency & ayou 9 0