SLIDE 1
Update Parallelism
April 30, 2018
1
HW 3 Posted
2
Update Parallelism April 30, 2018 1 HW 3 Posted 2 Parallelism - - PowerPoint PPT Presentation
Update Parallelism April 30, 2018 1 HW 3 Posted 2 Parallelism - - PowerPoint PPT Presentation
Update Parallelism April 30, 2018 1 HW 3 Posted 2 Parallelism Models Option 4: Shared Nothing in which all communication is explicit. CPU Memory Disk Well be talking about shared nothing today. Other models are easier
SLIDE 2
SLIDE 3
Parallelism Models
3
CPU Memory Disk …
We’ll be talking about “shared nothing” today. Other models are easier to work with.
Option 4: “Shared Nothing” in which all communication is explicit.
SLIDE 4
Data Parallelism
4
A A A C B A
Replication Partitioning
(needed for safety)
SLIDE 5
Updates
- Non-Serializable Schedules
- One Compute Node Fails
- A Communication Channel Fails
- Messages are
5
What can go wrong?
Node 1
T1: W(X) T2: W(X) T2: W(Y) T1: W(Y)
SLIDE 6
Updates
- Non-Serializable Schedules
- One Compute Node Fails
- A Communication Channel Fails
- Messages are
5
What can go wrong?
Node 1
T1: W(X) T2: W(X) T2: W(Y) T1: W(Y)
X
SLIDE 7
Updates (in Parallel)
- Non-Serializable Schedules
- One Compute Node Fails
- A Communication Channel Fails
- Messages delivered out-of-order
6
What can go wrong?
Node 1 Node 2
SLIDE 8
Updates (in Parallel)
- Non-Serializable Schedules
- One Compute Node Fails
- A Communication Channel Fails
- Messages delivered out-of-order
6
What can go wrong?
Node 1 Node 2
SLIDE 9
Updates (in Parallel)
- Non-Serializable Schedules
- One Compute Node Fails
- A Communication Channel Fails
- Messages delivered out-of-order
7
What can go wrong?
Node 1 Node 2
SLIDE 10
Updates (in Parallel)
- Non-Serializable Schedules
- One Compute Node Fails
- A Communication Channel Fails
- Messages delivered out-of-order
7
What can go wrong?
Node 1 Node 2
SLIDE 11
Updates (in Parallel)
- Non-Serializable Schedules
- One Compute Node Fails
- A Communication Channel Fails
- Messages delivered out-of-order
8
What can go wrong?
Node 1 Node 2
SLIDE 12
Updates (in Parallel)
- Non-Serializable Schedules
- One Compute Node Fails
- A Communication Channel Fails
- Messages delivered out-of-order
8
What can go wrong?
Node 1 Node 2
X Y Y X
SLIDE 13
Updates (in Parallel)
- Non-Serializable Schedules
- One Compute Node Fails
- A Communication Channel Fails
- Messages delivered out-of-order
9
What can go wrong?
SLIDE 14
Updates (in Parallel)
- Non-Serializable Schedules
- One Compute Node Fails
- A Communication Channel Fails
- Messages delivered out-of-order
9
What can go wrong? Classical Xacts
SLIDE 15
Updates (in Parallel)
- Non-Serializable Schedules
- One Compute Node Fails
- A Communication Channel Fails
- Messages delivered out-of-order
9
What can go wrong? Classical Xacts “Partitions”
SLIDE 16
Updates (in Parallel)
- Non-Serializable Schedules
- One Compute Node Fails
- A Communication Channel Fails
- Messages delivered out-of-order
9
What can go wrong? Classical Xacts “Partitions” Consensus
SLIDE 17
Data Parallelism
10
A A A C B A
Replication Partitioning
(needed for safety)
SLIDE 18
Simple Consensus
11
A A B B
Node 1 Node 2
Master Slave
Y X Y X
“Safe” … but Node 1 is a bottleneck.
SLIDE 19
Simpl-ish Consensus
12
A A
Node 1 Node 2
Master for A Master for B Node 2 agrees to Node 1’s order for A. Node 1 agrees to Node 2’s order for B.
B B
Y X Y X
SLIDE 20
Partitions
13
Node 1 Node 2 Node 1
From Node 1’s perspective, these are the same! Channel Failure Node Failure
Node 2
SLIDE 21
Failure Recovery
- Node Failure
- The node restarts and resumes serving requests.
- Channel Failure
- Node 1 and Node 2 regain connectivity.
14
SLIDE 22
Partitions
15
Node 1 Node 2
A=1 B=5 A=1 B=5
SLIDE 23
Partitions
16
Node 1
Option 1: Node 1 takes over
Node 2
A=1 B=5
SLIDE 24
Partitions
17
Node 1
Option 1: Node 1 takes over
Node 2
Node 2 is down. I control A & B now!
A=1 B=5
SLIDE 25
Partitions
18
Node 1
Option 1: Node 1 takes over
Node 2
Node 2 is down. I control A & B now! A = 2 B = 6
A=2 B=6
SLIDE 26
Partitions
19
Node 1
Option 1: Node 1 takes over
Node 2
A=2 B=6 A=2 B=6
SLIDE 27
Partitions
20
Node 1
Option 1: Node 1 takes over
Node 2
A=1 B=5 A=1 B=5
SLIDE 28
Partitions
21
Node 1
Option 1: Node 1 takes over
Node 2
A=1 B=5 A=1 B=5
Node 2 is down. I control A & B now!
SLIDE 29
Partitions
22
Node 1
Option 1: Node 1 takes over
Node 2
A=2 B=6 A=1 B=5
Node 2 is down. I control A & B now! A = 2 B = 6
SLIDE 30
Partitions
23
Node 1
Option 1: Node 1 takes over
Node 2
A=2 B=6 A=1 B=5 INCONSISTENCY!
SLIDE 31
Partitions
24
Node 1 Node 2
Option 2: Wait
SLIDE 32
Partitions
25
Node 1 Node 2
A = 2 B = 6
Option 2: Wait
SLIDE 33
Partitions
26
Node 1 Node 2
I can’t talk to Node 2 Let me wait! A = 2 B = 6
Option 2: Wait
SLIDE 34
Partitions
27
Node 1 Node 2
I can’t talk to Node 2 Let me wait! A = 2 B = 6
Option 2: Wait
SLIDE 35
Partitions
28
Node 1 Node 2
I can’t talk to Node 2 Let me wait! A = 2 B = 6 All set
Option 2: Wait
SLIDE 36
Partitions
29
Node 1 Node 2
Option 2: Wait
SLIDE 37
Partitions
29
Node 1
Option 2: Wait
SLIDE 38
Partitions
30
Node 1
I can’t talk to Node 2 Let me wait! A = 2 B = 6
Option 2: Wait
SLIDE 39
Partitions
31
Node 1
I can’t talk to Node 2 Let me wait! A = 2 B = 6 Still waiting…
Option 2: Wait
SLIDE 40
Partitions
32
Option 1: Assume Node Failure All data is available… but at risk of inconsistency. Option 2: Assume Connection Failure All data is consistent… but unavailable
SLIDE 41
33
C A P
- n
s i s t e n c y v a i l a b i l i t y
- r
- r
a r t i t i
- n
Traditionally: Pick any 2 T
- l
e r a n c e
SLIDE 42
34
C A P
- r
during
- n
s i s t e n c y v a i l a b i l i t y a r t i t i
- n
s I prefer this phrasing
SLIDE 43
Simpl-ish Consensus
35
A A
Node 1 Node 2
Master for A Master for B Node 2 agrees to Node 1’s order for A. Node 1 agrees to Node 2’s order for B.
B B
Y X Y X
SLIDE 44
Simpl-ish Consensus
36
A A
Node 1 Node 2
Master for A Master for B What if we need to coordinate between A & B?
B B
Y X Y X Withdraw $1000 from A Deposit $1000 into B
SLIDE 45
Naive Commit
37
Node 1 Node 2 Coordinator W(A,B)
SLIDE 46
Naive Commit
37
Node 1 Node 2 Coordinator Safe to Commit? W(A,B)
SLIDE 47
Naive Commit
37
Node 1 Node 2 Coordinator ACK Safe to Commit ? W(A,B)
SLIDE 48
Naive Commit
37
Node 1 Node 2 Coordinator ACK Safe to Commit W(A,B)
SLIDE 49
38
That packet sure does look tasty…
SLIDE 50
Naive Commit
39
Node 1 Node 2 Coordinator W(A,B) ACK
Is it safe to abort?
SLIDE 51
Naive Commit
40
Node 1 Node 2 Coordinator ACK ACK
What now?
W(A,B)
SLIDE 52
Naive Commit
41
Node 1 Node 2 Coordinator W(A) ACK
How do we know Node 2 even still exists?
SLIDE 53
2-Phase Commit
- One site selected as a coordinator.
- Initiates the 2-phase commit process.
- Remaining sites are subordinates.
- Only one coordinator per xact.
- Different xacts may have different coordinators.
42
SLIDE 54
Assumptions
- Undo/Redo Logging at Participants
- Participants can Abort an Xact at any time
- Participants can recover from a crash
- Redo Logging at Coordinator
- Coordinator can recover from a crash
- All logs replicated (to recover from hard failures)
43
SLIDE 55
Phase 1 - Prepare
44
Coordinator Node 1 Node 2
SLIDE 56
Phase 1 - Prepare
45
Coordinator Node 1 Node 2 “Prepare”
SLIDE 57
Phase 1 - Prepare
46
Coordinator Node 1 Node 2 “Prepare” “Commit” “Commit”
SLIDE 58
Phase 1 - Prepare
47
Coordinator Node 1 Node 2 “Prepare”
We are go for Commit
“Commit” “Commit”
SLIDE 59
Phase 2 - Commit
48
Coordinator Node 1 Node 2 “Prepare” “Commit”
We are go for Commit
“Commit” “Commit”
SLIDE 60
Phase 2 - Commit
49
Coordinator Node 1 Node 2 “Prepare” “Commit” “ACK” “ACK”
ACKs received Commit successful We are go for Commit
“Commit” “Commit”
SLIDE 61
Aborting
50
Coordinator Node 1 Node 2 “Prepare”
Commit Canceled
“Abort” “ACK” “ACK”
ACKs received Abort successful
“Commit” “Abort”
If any participant aborts in Phase 1, everyone aborts.
SLIDE 62
Guarantees
51
Coordinator Node 1 Node 2 “Prepare” “Commit” “ACK” “ACK”
ACKs received Commit successful We are go for Commit
“Commit” “Commit”
A Node “Commit” means the node is able to commit. A Coordinator “Commit” means the transaction must commit.
SLIDE 63
Guarantees
52
Coordinator Node 1 Node 2 “Prepare” “Commit” “ACK” “ACK”
ACKs received Commit successful We are go for Commit
“Commit” “Commit”
Once a node commits, the xact is still not committed yet. However the node must avoid breaking the commit.
SLIDE 64
Failure Modes
53
Coordinator Node 1 Node 2 “Prepare” “Commit”
SLIDE 65
Failure Modes
53
Coordinator Node 1 Node 2 “Prepare” “Commit”
Prepare unreceived and unacknowledged: Coordinator (1) Retries, or (2) Aborts
SLIDE 66
Failure Modes
54
Coordinator Node 1 Node 2 “Prepare” “Commit”
CRASH!
SLIDE 67
Failure Modes
54
Coordinator Node 1 Node 2 “Prepare” “Commit”
Node 2 crashes before responding: Restart and continue as a dropped packet
CRASH!
SLIDE 68
Failure Modes
55
Coordinator Node 1 Node 2 “Prepare” “Commit” “Commit”
SLIDE 69
Failure Modes
55
Coordinator Node 1 Node 2 “Prepare” “Commit” “Commit”
Node “Commit” unreceived: (1) Re-sent “Prepare” can be ignored. (2) Node still able to abort.
SLIDE 70
Failure Modes
56
Coordinator Node 1 Node 2 “Prepare” “Commit” “Commit”
CRASH!
SLIDE 71
Failure Modes
56
Coordinator Node 1 Node 2 “Prepare” “Commit” “Commit”
Node 2 crashes after responding: Restart from log
CRASH!
SLIDE 72
Failure Cases
57
Coordinator Node 1 Node 2 “Prepare” “Commit” “ACK”
We are go for Commit
“Commit” “Commit”
SLIDE 73
Failure Cases
57
Coordinator Node 1 Node 2 “Prepare” “Commit” “ACK”
We are go for Commit
“Commit” “Commit”
Coordinator “Commit” unreceived: Commit must happen, coordinator resends
SLIDE 74
Failure Cases
58
Coordinator Node 1 Node 2 “Prepare” “Commit” “ACK”
We are go for Commit
“Commit” “Commit”
CRASH!
SLIDE 75
Failure Cases
58
Coordinator Node 1 Node 2 “Prepare” “Commit” “ACK”
We are go for Commit
“Commit” “Commit”
Node 2 crash: Restart. Already logged “Commit” message, so all is well.
CRASH!
SLIDE 76
Failure Cases
59
Coordinator Node 1 Node 2 “Prepare” “Commit” “ACK” “ACK”
We are go for Commit
“Commit” “Commit”
SLIDE 77
Failure Cases
59
Coordinator Node 1 Node 2 “Prepare” “Commit” “ACK” “ACK”
We are go for Commit
“Commit” “Commit”
Node “Ack” unreceived: Ok. Resent “Commit” ignored by node
SLIDE 78
Failure Cases
60
Coordinator Node 1 Node 2 “Prepare” “Commit” “ACK”
We are go for Commit
“Commit” “Commit”
CRASH!
“ACK”
SLIDE 79
Failure Cases
60
Coordinator Node 1 Node 2 “Prepare” “Commit” “ACK”
We are go for Commit
“Commit” “Commit”
Node crash after “Ack”: Ok. Log already recorded commit
CRASH!
“ACK”
SLIDE 80
Replication
- Mode 1: Periodic Backups
- Copy the replicated data nightly/hourly.
- Mode 2: Log Shipping
- Only send changes (replica serves as the log).
61
SLIDE 81
Replication
- Mode 1: Periodic Backups
- Copy the replicated data nightly/hourly.
- Mode 2: Log Shipping
- Only send changes (replica serves as the log).
61
SLIDE 82
Replication
- Ensuring durability
- Ensuring write-consistency under 2PC
- Ensuring read-consistency without 2PC
62
SLIDE 83
Ensuring Durability
63
When is a replica write durable?
SLIDE 84
Ensuring Durability
64
Never.
SLIDE 85
Ensuring Durability
64
Never. What you should be asking is how much durability do you need?
SLIDE 86
Ensuring Durability
65
For N Failures N+1 Replicas (Assuming Failure = Crash)
SLIDE 87
Ensuring Write Consistency
66
Coordinator Node 1 “Prepare” “Commit” Node 1 asserts that the commit is durable! What if Node 1 fails?
SLIDE 88
Ensuring Write Consistency
67
Coordinator Node 1 Replica “Prepare” “Commit” “Prepare” “Commit”
SLIDE 89
Ensuring Write Consistency
67
Coordinator Node 1 Replica “Prepare” “Commit” “Prepare” “Commit” Waiting for Node 1 to replicate is slooooow! Let the coordinator take over!
SLIDE 90
Ensuring Write Consistency
68
Coordinator Node 1 Replica “Prepare” “Commit” “Commit”
SLIDE 91
Ensuring Write Consistency
68
Coordinator Node 1 Replica “Prepare” Like 2PC… … but better. We may not need to wait for the replica “Commit” “Commit”
SLIDE 92
Ensuring Write-Consistency
69
Replica 2 Replica 1 Replica 3
A: Prepare
Coordinator Alice Coordinator Bob
B: Prepare A: Prepare B: Prepare A: Prepare B: Prepare
SLIDE 93
Ensuring Write-Consistency
70
Replica 2 Replica 1 Replica 3
Coordinator Alice Coordinator Bob
A: Prepare B: Prepare A: Prepare B: Prepare A: Prepare B: Prepare
SLIDE 94
Ensuring Write-Consistency
71
Replica 2 Replica 1 Replica 3
Coordinator Alice Coordinator Bob
B: Prepare B: Prepare A: Prepare
Commit! Commit!
SLIDE 95
Ensuring Write-Consistency
72
Majority Vote N Replicas (N/2)+1 Votes Needed
SLIDE 96
Ensuring Read Consistency
73
Forget transactions, let’s go back to reads & writes Can we do better than 2PC if we don’t need xacts?
SLIDE 97
74
Replica 2 Replica 1 Replica 3
W(A = 3) (1) Alice writes ‘A’
SLIDE 98
75
Replica 2 Replica 1 Replica 3
W(A = 3) (1) Alice writes ‘A’ (2) Alice tells Bob
SLIDE 99
76
Replica 2 Replica 1 Replica 3
W(A = 3) (1) Alice writes ‘A’ (2) Alice tells Bob (3) Bob reads ‘A’ R(A)
SLIDE 100
77
Replica 2 Replica 1 Replica 3
W(A = 42) (1) Alice writes ‘A’ (2) Alice tells Bob (3) Bob reads ‘A’ R(A) What can we do to guarantee that Bob will see the 42?
SLIDE 101
Ensuring Read Consistency
78
Approach: Alice and Bob each wait for multiple responses. Alice waits for ‘ack’s Bob waits for read responses. How many responses are required for each?
SLIDE 102
79
Replica 2 Replica 1 Replica 3
W(A = 42) R(A) ACK
SLIDE 103
80
Replica 2 Replica 1 Replica 3
W(A = 42) R(A) ACK “666”
SLIDE 104
81
Replica 2 Replica 1 Replica 3
W(A = 42) R(A) ACK “666”
SLIDE 105
82
Replica 2 Replica 1 Replica 3
W(A = 42) R(A) ACK “666”
SLIDE 106
Ensuring Read-Consistency
83
Like Majority Vote N Replicas R Replica Reads Needed W Writer Acks Needed R + W > N