The SMART Way to Migrate Replicated Stateful Services
Jacob R. Lorch, Atul Adya, Bill Bolosky, Ronnie Chaiken, John Douceur, Jon Howell
Microsoft Research
First EuroSys Conference19 April 2006
The SMART Way to Migrate Replicated Stateful Services
Paxos
Replicated
BCA
services
statefulReplicated stateful
services
• Problem: Machine failure leads to unavailability– Solution: Replicate the service for fault tolerance
• Problem: Replica state can become inconsistent– Solution: Use replicated state machine approach
The SMART Way to Migrate Replicated Stateful Services
Migrating replicated services
B CA D E
• Migration: Changing the configuration – the set of machines running replicas
• Uses of migration– Replace failed machines for long-term fault tolerance– Load balancing– Increasing or decreasing number of replicas
The SMART Way to Migrate Replicated Stateful Services
• Can remove non-failed machines– Enables autonomic migration, i.e., migration without human
involvement– Enables load balancing
• Can do concurrent request processing• Can perform arbitrary migrations, even ones replacing
entire configuration• Completely described in our paper
Limitations of current approaches
• Cannot remove non-failed machines without creating window of vulnerability– Can only remove known-failed machines– Cannot use migration for load balancing
• Cannot process requests in parallel
Limitations of current approaches
addressed by SMART
The SMART Way to Migrate Replicated Stateful Services
Outline
• Introduction• Background on Paxos• Limitations of existing approaches• SMART: Service Migration And Replication
Technique– Configuration-specific replicas– Shared execution modules
• Implementation and evaluation• Conclusions
The SMART Way to Migrate Replicated Stateful Services
Background on Paxos
The SMART Way to Migrate Replicated Stateful Services
Paxosprotocol
Background: Paxos overview
• Goal: Every service replica runs the same sequence of requests– Deterministic service ensures state changes and replies are
consistent• Approach: Paxos assigns requests to virtual “slots”
– No two replicas assign different requests to same slot– Each replica executes requests in slot order
BA C
1 2 3 4 5 6…
slots:requests:
1 2 3 4 5 6 1 2 3 4 5 6… …
The SMART Way to Migrate Replicated Stateful Services
Background: Paxos protocol
• One replica is the leader• Clients send requests to the leader• Leader proposes a request by sending PROPOSE
message to all replicas• Each replica logs it and sends a LOGGED message
to the leader• When leader receives LOGGED messages from a
majority, it decides it and sends a DECIDED message
BA CZ
client server replicas
ReqLOGGED LOGGEDPROPOSEPROPOSEDECIDEDDECIDED
The SMART Way to Migrate Replicated Stateful Services
Background: Paxos leader change
• If leader fails, another replica “elects” itself• New leader must poll replicas and hear
replies from a majority– Ensures it learns enough about previous
leaders’ actions to avoid conflicting proposals
A B CPollPoll Reply
The SMART Way to Migrate Replicated Stateful Services
Background: Paxos migration
• Service state includes current configuration– Request that changes that part of the state migrates the
service
• Configuration after request n responsible for requests n+α and beyond
A
B
C 79
79
79Servicestate
A, B, DA, B, C80
80
80α
81
81
81
82
82
82
83
83
83
84
84
84
D 85
85
85
The SMART Way to Migrate Replicated Stateful Services
Rationale for α
• With α=1, slot n can change the configuration responsible for slot n+1
• Leader can’t propose slot n+1 until n is decided– Doesn’t know who to make proposal to, let alone whether it
can make proposal at all
• Prevents pipelining of requests– Request may wait a network round trip and a disk write
BA CZReqReq PROPOSEPROPOSE LOGGED LOGGED
The SMART Way to Migrate Replicated Stateful Services
Limitations of existing approaches
The SMART Way to Migrate Replicated Stateful Services
No request pipelining
• Leader change is complicated– How to ensure that new leader knows the right
configuration to poll?– How to handle some outstanding proposals being
from one configuration and some from another?– Other problems
• To avoid this complexity, current approaches use α=1
• But, this prevents request pipelining
The SMART Way to Migrate Replicated Stateful Services
Window of vulnerability
• Removing a machine creates window of vulnerability– Effectively, it induces a failure of the removed replica– Consequently, service can become permanently unavailable
even if less than half the machines fail
• Considered acceptable since machines only removed when known to a human to have permanently failed
• Not suitable for autonomic migration using imperfect failure detectors, or for load balancing
BA C DDECIDEDPROPOSEPROPOSEDECIDED LOGGED LOGGEDPollPoll
The SMART Way to Migrate Replicated Stateful Services
SMART
The SMART Way to Migrate Replicated Stateful Services
Configuration-specific replicas
• Each configuration has its own set of replicas and its own separate instance of Paxos
• Simplifies leader change so we can pipeline requests– Election always happens in a static configuration
• No window of vulnerability because a replica can remain alive until next configuration is established
A
Replica 1B
B
Replica 2A Replica 2B
Replica 1C
C D
Replica 2D
Replica 1A
The SMART Way to Migrate Replicated Stateful Services
SMART migration protocol
• After creating new configuration, send JOIN msgs• After executing request n+α-1, send FINISHED msgs
– Tells new replicas where they can get starting state– Makes up for possibly lost JOIN messages
• When a majority of successor configuration have their starting state, replica kills itself
• If a machine misses this phase, it can still join later
A
Replica 1B
B
Replica 2A Replica 2B
Replica 1C
C D
Replica 2D
Replica 1AJOINJOINJOINFINISHEDFINISHEDFINISHED FINISHEDFINISHEDFINISHED FINISHEDFINISHEDFINISHEDFINISHEDFINISHEDFINISHED FINISHEDFINISHEDFINISHED FINISHEDFINISHEDFINISHED
READYREADYREADY READYREADYREADYPREPARE
JOIN-REQ
JOIN
The SMART Way to Migrate Replicated Stateful Services
Agreement 1A
Agreement 2A
Agreement 1B
Agreement 2B
Agreement 1C
Agreement 2D
Shared execution modules
• Configuration-specific replicas have a downside– One copy of service state for each replica– Need to copy state to new replicas
• Solution: Shared execution modules– Divide replica into agreement and execution modules– One execution module for all replicas on machine
Replica 1A
A
Replica 1B
B
Replica 2A Replica 2B
Replica 1C
C D
Replica 2D
Execution 1A
Execution 2A
Execution 1B
Execution 2B
Execution 1C
Execution 2DExecution Execution Execution Execution
The SMART Way to Migrate Replicated Stateful Services
Implementation and evaluation
• SMART implemented in a replicated state machine library, LibSMART– Lets you build a service as if it were single-machine, then
turns it into a replicated, migratable service
• Farsite distributed file system service ported to LibSMART– Straightforward because LibSMART uses BFT interface
• Experimental results using simple key/value service– Pipelining reduces average client latency by 14%– Migration happens quickly, so clients only see a bit of extra
latency, less than 30 ms
The SMART Way to Migrate Replicated Stateful Services
Conclusions
• Migration is useful for replicated services– Long-term fault tolerance, load balancing
• Current approaches to migration have limitations• SMART removes these limitations by using
configuration-specific replicas– Can remove live machines, enabling autonomic migration
and load balancing– Can overlap processing of concurrent requests
• SMART is practical– Implementation supports large, complex file system service