Analyzing Interactions of Asynchronously Communicating
Systems
Tevfik BultanDepartment of Computer Science
University of California, Santa [email protected]
http://www.cs.ucsb.edu/~bultan
Acknowledgements
• Joint work with– Xiang Fu, Hofstra University– Jianwen Su, University of California, Santa Barbara– Zachary Stengel, Microsoft– Samik Basu, Iowa State
Motivation 1: Web Services
• Web services support basic client/server style interactions
• Example: Amazon E-Commerce Web Service (AWS-ECS) • AWS-ECS WSDL specification lists 40 operations that provide differing
ways of browsing Amazon’s product database such as– ItemSearch, CartCreate, CartAdd, CartModify, CartGet, CartClear
• Based on the AWS-ECS WSDL specification one can implement clients that interact with AWS-ECS
Service Requester
Service Provider
Request
Response
SOAP
WSDL
Client Server
Composing Services
• Can this framework support more than basic client/server style interactions?
• Can we compose a set of services to construct a new service?
• For example:– If we are building a bookstore service, we may want to use both
Amazon’s service and Barnes & Noble’s service in order to get better prices
• Another (well-known) example:– A travel agency service that uses other services (such as flight
reservation, hotel reservation, and car rental services) to help customers book their trips
Orchestration vs Choreography
Orchestration: Define an executable process that interacts with existing services and executes them in a particular order and combines the results to achieve a new goal– From atomic services to stateful services– Web Services Business Process Execution Language (WS-BPEL)
Choreography: Specify how the individual services should interact with each other. Find or construct individual services that follow this interaction specification– Global specification of interactions among services– Web Services Choreography Description Language (WS-CDL)
A choreography can be realized by writing an orchestration for each peer involved in the choreography– Choreography as global behavior specification– Orchestration as local behavior specification that realizes the
global specification
Web Services Standards Stack
Data
Type
Service Orchestration
Protocol
Web Services Business Process Execution Language (WS-BPEL)
Web Services Description Language (WSDL)
Simple Object Access Protocol (SOAP)
XML Schema (XSD)
Extensible Markup Language (XML)
AtomicService
AtomicService
OrchestratedService
SOAP
WSDL
WSDL
Choreography Web Services Choreography Description Language (WS-CDL)
WS-BPEL
OrchestratedService
WS-BPEL
SOAP
SOAP
SOAP
SOAPWS-CDL
Asynchronous Messaging
• Sender does not have to wait for the receiver– Message is inserted to a message queue– Messaging platform guarantees the delivery of the message
• Why support asynchronous messaging?– Otherwise the sender has to block and wait for the receiver – Sender may not need any data to be returned– If the sender needs some data to be returned, it should only wait
when it needs to use that data– Asynchronous messaging can alleviate the latency of message
transmission through the Internet– Asynchronous messaging can prevent sender from blocking if the
receiver service is temporarily unavailable• Rather then creating a thread to handle the send, use
asynchronous messaging
Motivation 2: Singularity OS
• Experimental OS developed by Microsoft Research to explore new ideas for operating system design
• Key design principles: – Dependability– Security
• Key architectural decision:– Implement a sealed process system
• Software Isolated Processes (SIPs)– Closed code space (no dynamic code loading or code generation)– Closed object space (no shared memory)
• Inter-process communication occurs via message passing over channels
Singularity Channels
• Channels allow 2-Party asynchronous communication via FIFO message queues– Sends are non blocking– Receives block until a message is at the head of a receive queue
• Each channel has exactly two endpoints– Type exposed for each endpoint (Exp and Imp)– Each endpoint owned by at most one process at any time
• Owner of Exp referred to as Server• Owner of Imp referred to as Client
Channel Contracts
• Written in Sing #• Contracts specify two things:
1. The messages that may be sent over a channel• out message are sent from the
Server endpoint to the Client endpoint (SC)
• in messages are sent from the Client endpoint to the Server endpoint (CS)
– The set of allowed message sequences• out message marked with !• in messages marked with ?
public contract KeyboardDeviceContract { out message AckKey( uint key ); out message NakKey(); out message Success();
in message GetKey(); in message PollKey();
state Start { Success! -> Ready; }
state Ready { GetKey? -> Waiting; PollKey? -> (AckKey! or NakKey!) -> Ready; }
state Waiting { AckKey! -> Ready; NakKey! -> Ready; }}
public contract KeyboardDeviceContract { out message AckKey( uint key ); out message NakKey(); out message Success();
in message GetKey(); in message PollKey();
state Start { Success! -> Ready; }
state Ready { GetKey? -> Waiting; PollKey? -> (AckKey! or NakKey!) -> Ready; }
state Waiting { AckKey! -> Ready; NakKey! -> Ready; }}
• A contract specifies a finite state machine• Each message causes a deterministic transition from one state to
another state
KeyboardDeviceContract
Channel Contracts
Start
Ready$0ReadyWaiting
SC:Success
SC:AckKey
SC:AckKey
CS:GetKey
CS:PollKey
SC:NakKey
SC:AckKey
Implicit State
Outline
• Motivation – Composition of Web Services – Singularity Channel Contracts
• Conversations• Realizability• Synchronizability • Applications• Recent Results
Going to Lunch at UCSB
• Before Xiang left UCSB, Xiang, Jianwen and I were using the following protocol for going to lunch:
– Sometime around noon one of us would call another one by phone and tell him where and when we would meet for lunch.
– The receiver of this first call would call the remaining peer and pass the information.
• Let’s call this protocol the First Caller Decides (FCD) protocol.
• At the time we did not have answering machines or voicemail!
FCD Protocol Scenarios
• Possible scenario
1. Tevfik calls Jianwen with the decision of where and when to eat
2. Jianwen calls Xiang and passes the information• Another scenario
1. Jianwen calls Tevfik with the decision of where and when to eat
2. Tevfik calls Xiang and passes the information• Yet another scenario
1. Tevfik calls Xiang with the decision of where and when to eat• Maybe Jianwen also calls Xiang at the same time with a
different decision. But the phone is busy.• Jianwen keeps calling. But Xiang is not going to answer
because according to the protocol the next thing Xiang has to do is to call Jianwen.
2. Xiang calls Jianwen and passes the information
FCD Protocol: Tevfik’s Behavior
Tevfik calls Jianwen with the lunch decision
Let’s look at all possible behaviors of Tevfik based on the FCD protocol
Tevfik is hungry
Tevfik calls Xiang with the lunch decision
Tevfik receives a call from Jianwen passing him the
lunch decision
Tevfik receives a call from Xiang passing him the
lunch decisionTevfik receives a call from Xiang telling him the lunch decision that Tevfik has to pass to
Jianwen
FCD Protocol: Tevfik’s Behavior
!T->J:D
!T->X:D
?J->T:P
?X->T:P
?X->T:D?J->T:D
!T->J:P!T->X:P
T->J:D Tevfik calls Jianwen with the lunch decision
Message Labels:
! send
? receiveJ->X:P
Jianwen calls Xiang to pass the decision
!T->J:D
?X->T:D
!T->J:D
?J->T:D
!T->X:P
Tevfik
!T->J:P
?J->T:P
?X->T:P
!X->J:D
?T->X:D
!X->T:D
?J->X:D
!X->T:P
Xiang
!X->J:P
?J->X:P
?T->X:P
!J->T:D
?X->J:D
!J->X:D
?T->J:D
!J->X:P
Jianwen
!J->T:P
?T->J:P
?X->J:P
State machines for the FCD Protocol
• Three state machines characterizing the behaviors of Tevfik, Xiang and Jianwen according to the FCD protocol
FCD Protocol Has Voicemail Problems
• When the university installed a voicemail system FCD protocol started causing problems– We were showing up at different restaurants at different times!
• Example scenario: – Tevfik calls Xiang with the lunch decision – Jianwen also calls Xiang with the lunch decision
• The phone is busy (Xiang is talking to Tevfik) so Jianwen leaves a message
– Xiang calls Jianwen passing the lunch decision• Jianwen does not answer (he already left for lunch) so Xiang
leaves a message– Jianwen shows up at a different restaurant!
• Message sequence is: T->X:D J->X:D X->J:P– The messages J->X:D and X->J:P are never consumed
• This scenario is not possible without voicemail!
A Different Lunch Protocol
• To fix this problem, Jianwen suggested that we change our lunch protocol as follows:
– As the most senior researcher among us Jianwen would make the first call to either Xiang or Tevfik and tell when and where we would meet for lunch.
– Then, the receiver of this call would pass the information to the other peer.
• Let’s call this protocol the Jianwen Decides (JD) protocol
?X->T:P?J->T:D
!T->X:P
Tevfik XiangJianwen
?T->X:P?J->X:D
!X->T:P
!J->T:D !J->X:D
State machines for the JD Protocol
• JD protocol works fine with voicemail!
Conversations
• The FCD and JD protocols specify a set of conversations – A conversation is the sequence of messages generated during an
execution of the protocol
• We can specify the set of conversations without showing how the peers implement them– we call such a specification a conversation protocol
T->J:D
T->X:D
X->J:P
X->T:D X->J:DJ->T:D
J->X:D
J->X:PT->J:P J->T:P
T->X:P
X->T:P
FCD Protocol
J->T:D J->X:D
T->X:P X->T:P
JD Protocol
FCD and JD Conversation Protocols
Conversation set: { T->X:D X->J:P, T->J:D J->X:P, X->T:D T->J:P, X->J:D J->T:P, J->T:D T->X:P,
J->X:D X->T:P }
Conversation set: { J->T:D T->X:P,
J->X:D X->T:P}
Observations & Questions
• The implementation of the FCD protocol behaves differently with synchronous and asynchronous communication whereas the implementation of the JD protocol behaves the same. – Can we find a way to identify such implementations?
• The implementation of the FCD protocol does not obey the FCD protocol if asynchronous communication is used whereas the implementation of the JD protocol obeys the JD protocol even if asynchronous communication used.– Given a conversation protocol can we figure out if there is an
implementation which generates the same conversation set?
Conversations, Choreography, Orchestration
• Peer state machines are orchestrations– A peer state machine can be specified using an orchestration
language such as WS-BPEL– One can translate WS-BPEL specifications to peer state machines
• A conversation protocol is a choreography specification– A conversation set corresponds to a choreography– A conversation set can be specified using a choreography
language such as WS-CDL– One can translate WS-CDL specifications to conversation protocols
Bottom-Up vs. Top-Down
Bottom-up approach• Specify the behavior of each peer
– For example using an orchestration language such as WS-BPEL• The global communication behavior (conversation set) is implicitly
defined based on the composed behavior of the peers• Global communication behavior is hard to understand and analyze
Top-down approach• Specify the global communication behavior (conversation set) explicitly
as a protocol– For example using a choreography language such as WS-CDL
• Ensure that the conversations generated by the peers obey the protocol
ConversationProtocol(ChoreographySpecification)
GF(T->X(P) X->T(P))? LTL property
InputQueue
...Virtual Watcher?
LTL property
Peer T Peer XPeer J
J->T:D J->X:D
T->X:P X->T:P
GF(T->X(P) X->T(P))
!J->T:D
!J->X:D
?X->T:P
?J->T:D
!T->X:P
?T->X:P?J->X:D
!X->T:P
Top-Down vs. Bottom-Up
Outline
• Motivation – Composition of Web Services – Singularity Channel Contracts
• Conversations• Realizability• Synchronizability • Applications• Recent Results
Realizability Question
• Conversation protocol specifies the global communication behavior– How do we implement the peers?
• How do we obtain the contracts that peers have to obey from the global contract specified by the conversation protocol?
– Synthesize peer implementations by projecting the global protocol to each peer by dropping unrelated messages for each peer
If this equality holds the conversation protocol is realizable• The JD protocol is realizable • The FCD protocol is not realizableAre there conditions which ensure the equivalence?
Conversations generated by the projected services
Conversations specified by the conversation protocol
?
Realizability Problem
• Not all conversation protocols are realizable!
AB: m1
CD: m2
Conversation protocol
Conversation “m2 m1m2 m1” will also be generated by all peer implementations which follow the protocol
!m1 ?m1 !m2 ?m2
Peer A Peer B Peer C Peer D
Projection of the conversation protocol to the peers
Realizability Conditions
Three sufficient conditions for realizability (no message content) • Lossless join
– Conversation set should be equivalent to the join of its projections to each peer
• Synchronous compatible– When the projections are composed synchronously, there should
not be a state where a peer is ready to send a message while the corresponding receiver is not ready to receive
• Autonomous– At any state, each peer should be able to do only one of the
following: send, receive or terminate
(a peer can still choose among multiple messages)
Realizability Conditions
AB: m1
CD: m2
AB: m1BA: m2
AC: m3
BA: m2
AB: m1
• Following protocols fail one of the three conditions but satisfy the other two
Not lossless join
Not autonomous
AB: m1
CA: m2
Not synchronous compatible
Outline
• Motivation – Composition of Web Services – Singularity Channel Contracts
• Conversations• Realizability• Synchronizability • Some Experiments• Applications
Bottom-Up Approach
• We know that analyzing conversations of composite web services is difficult due to asynchronous communication– Model checking for conversation properties is undecidable even for
finite state peers
• The question is:– Can we identify the composite web services where asynchronous
communication does not create a problem?• We call such compositions synchronizable
• The implementation of the JD protocol is synchronizable • The implementation of the FCD protocol is not synchronizable
Three Examples, Example 1
requester server
!r2
?a1 ?a2
!e
!r1
• Conversation set is regular: (Conversation set is regular: (rr11aa11 | | rr22aa22)* )* ee
• During all executions the message queues are bounded
r1, r2
a1, a2
e ?r1
!a1 !a2
?r2
?e
Example 2
• Conversation set is not regularConversation set is not regular• Queues are not bounded
requester server
!r2
?a1 ?a2
!e
!r1
r1, r2
a1, a2
e ?r1
!a1 !a2
?r2
?e
Example 3
• Conversation set is regular: (Conversation set is regular: (rr11 | | rr22 | | rara)* )* ee
• Queues are not bounded
requester server
!r2
?a !r
!e!r1
r1, r2
a1, a2
e
?r1 ?r2
?e
?r !a
State Spaces of the Three Examples
0
200
400
600
800
1000
1200
1400
1600
1 3 5 7 9 11 13
Example 1
Example 2
Example 3
queue length
# o
f st
ates
in
th
ou
san
ds
• Verification of Examples 2 and 3 are difficult even if we bound the queue length
• How can we distinguish Examples 1 and 3 (with regular conversation sets) from 2?– Synchronizability Analysis
Synchronizability Analysis
• A composite web service is synchronizable if its conversation set does not change – when asynchronous communication is replaced with synchronous
communication
• If a composite web service is synchronizable we can check the properties about its conversations using synchronous communication semantics – For finite state peers this is a finite state model checking problem
Synchronizability Analysis
Sufficient conditions for synchronizability:• A composite web service is synchronizable, if it satisfies the
synchronous compatible and autonomous conditions
• Connection between realizability and synchronizability:– A conversation protocol is realizable if its projections to peers are
synchronizable and the protocol itself satisfies the lossless join condition
Outline
• Motivation – Composition of Web Services – Singularity Channel Contracts
• Conversations• Realizability• Synchronizability • Applications• Recent Results
Are These Conditions Too Restrictive?
Problem Set Size Pass?Source Name #msg #states #trans.
ISSTA’04 SAS 9 12 15 yes
IBM
Conv.
Support
Project
CvSetup 4 4 4 yesMetaConv 4 4 6 no
Chat 2 4 5 yesBuy 5 5 6 yes
Haggle 8 5 8 noAMAB 8 10 15 yes
BPEL
spec
shipping 2 3 3 yes
Loan 6 6 6 yes
Auction 9 9 10 yesCollaxa.
com
StarLoan 6 7 7 yesCauction 5 7 6 yes
Singularity Channel Contract Verification
• State machine construction allows for automated verification and analysis of channel communication
• Singularity compiler automatically checks compliance of client and server processes to the specified contract
• Claim from Singularity documentation:– "clients and servers that have been verified separately against the
same contract C are guaranteed not to deadlock when allowed to communicate according to C.“
• This claim is wrong!
IO_RUNNING$0ReadyState$1
ReadyState$0
IO_RUNNING
Deadlock Example: The TpmContract
IO_RUNNING$0
IO_RUNNINGReadyState
ReadyState$1
ReadyState$0
ReadyState
Server Projection
ClientProjection
CS:Send SC:AckStartSend
SC:SendComplete
CS:GetTpmStatusCS:GetTpmStatusSC:TpmStatus SC:TpmStatus
Send?
Send!
AckStartSend!
AckStartSend?
SendComplete!
GetTpmStatus!
GetTpmStatus?TpmStatus!
GetTpmStatus!TpmStatus? TpmStatus?
TpmStatus!GetTpmStatus?
ServerReceive Queue
ClientReceive Queue
Conversation
SendComplete?
Deadlock Example: The TpmContract
IO_RUNNING$0
IO_RUNNINGReadyState
ReadyState$1
ReadyState$0
Send?
IO_RUNNING$0
IO_RUNNINGReadyState
ReadyState$1
ReadyState$0
Server Projection
Send!
AckStartSend!
AckStartSend?
SendComplete!
GetTpmStatus!
GetTpmStatus?TpmStatus!
GetTpmStatus!TpmStatus? TpmStatus?
SendComplete?
TpmStatus!GetTpmStatus?
ServerReceive Queue
ClientReceive Queue
Send
AckStartSendSendComplete
GetTpmStatus
TpmStatus
ConversationCS: SendSC: AckStartSendSC: SendCompleteCS: GetTpmStatusSC: TpmStatus
ClientProjection
Realizability Problem
• KeyboardDeviceContract is not realizable – It violates the autonomous condition
• It turns out that autonomous condition is sufficient (but not necessary) for realizability of two-party protocols (Singularity channel contracts are two-party protocols)– If a contract is autonomous, it is guaranteed to be realizable– However, it can be realizable but not autonomous
• i.e., false positives are possible when we use autonomous condition as our realizability check
Autonomous condition and false positives
• Example: TpmContract
IO_RUNNING$0
IO_RUNNINGReadyState
ReadyState$1
ReadyState$0
CS:Send SC:AckStartSend
SC:SendComplete
CS:GetTpmStatusCS:GetTpmStatusSC:TpmStatus SC:TpmStatus
IO_RUNNING$1
SC:SendCompleteSC:TpmStatus
Fixed
Violates Autonomous
condition
• Explicit state verification is expensive using asynchronous communication– Exponential state space explosion in the worst case
• Example: BlowupKContract
Model checking efficiency
S1
S2
…
Sk
SC:m1
SC:m1
SC:m1
SC:m2
SC:m2
SC:m2
CS:m3
Model checking efficiency
• If contract is realizable, conversations generated using asynchronous communication and synchronous communication are the same– Therefore, synchronous communication model can be used for
verification
S1
S2
…
Sk
SC:m1
SC:m1
SC:m1
SC:m2
SC:m2
SC:m2
CS:m3
Yes
Tune: A Tool For Analyzing Sing# Contracts
ChannelContract Contract
Parser
Contract State
Machine
ContractAnalyzer
Realizable?Synchronous
Promela Generator
Asynchronous Promela
Generator
Sync Promela
Async PromelaNo
Spin
Data Collector
LTL Formulas
Report
Consumed by
Produces
File
TuneComponentExternal Tool
LTL Formulas
Analysis Efficiency
• Performed autonomous check and exhaustive deadlock search for ~95% of contracts to compare analysis time
• Results show clear advantage to performing the autonomous check
LTL Property Validation
• Selected 10 contracts for LTL property validation
• Both synchronous and asynchronous models were used to compare performance
Realizability Results
• Ran analysis on 93 contracts from the Singularity code base (version 2.0) and documentation
• Found two contracts that violate the autonomous condition (TpmContract and ReservationSession)
• Exhaustive search showed deadlocking execution traces for both contracts– Confirmed by Singularity developers
• Tune did not report any false positives from autonomous check for any of the contracts analyzed
• In practice, autonomous condition is not too restrictive
Outline
• Motivation – Composition of Web Services – Singularity Channel Contracts
• Conversations• Realizability• Synchronizability • Applications• Recent Results
Recent Results
• Open problems (until recently):– Is realizability decidable?– Is synchronizability decidable?
• Recent result– Synchronizability is decidable!
• We are pretty sure that we will also be able to show that realizability is decidable
Synchronizability Result
• Given a set of peers– Let C-A be their asynchronous composition– Let C-k be their bounded-asynchronous composition where queues
are bounded to be of size k• A send to a full queue (i.e., a queue with k elements) blocks
– C-0 corresponds to synchronous composition
Theorem: A composition is synchronizable (i.e., L(C-0)=L(C-A)) if and only if the conversation set of C-0 and C-1 are the same (i.e., L(C-0) = L(C-1))
Reachability & Synchronizability
• It is well-known that reachability problems for asynchronously communicating systems are undecidable
• We can extend the synchronizability definition to include reachability – We call a composition reachability-synchronizable if 1) it is
synchronizable and 2) the set of states reachable in the synchronous composition is same as the set of empty-queue states reachable in the asynchronous composition
• Very recent result: Determining if a composition is reachability-synchronizable is decidable– Theorem: A composition is reachability-synchronizable if and only
if the conversation set and the (empty-queue) reachable states of C-0 and C-1 are the same
Related Work
• Singularity:– [Hunt, Larus SIGOPS ‘07] Singularity: rethinking the software stack– [Fähndrich, Aiken, Hawblitzel, et. al SIGOPS/Eurosys ‘07]
Language support for fast and reliable message-based communication in singularity os.
– Influenced by work on Session Types• [Honda, Vasconcelos, Kubo ESOP ’98] Language primitives
and type discipline for structured communication-based programming
– Source code and RDK: http://codeplex.com/singularity
Related Work
• Realizability of message sequence charts– [Alur, Etessami, Yannakakis] ICSE’00, ICALP’01] – Defines similar notion of realizability– Different conversation model
Related Work
• Specification approaches that are similar to conversation protocols
– [Parunak ICMAS 96] Visualizing agent conversations: Using enhanced Dooley graphs for agent design and analysis.
– [Hanson, Nandi, Kumaran EDOCC’02] Conversation support for business process integration
Related Work
• Message Sequence Charts (MSC)
– [Alur, Etassami, Yannakakis ICSE’00, ICALP’01] Realizability of MSCs and MSC Graphs
– [Uchitel, Kramer, Magee ACM TOSEM 04] Implied Scenarios in MSCs
Related Work
• Verification of web services– Petri Nets
• [Narayanan, McIlraith WWW’02] Simulation, verification, composition of web services using a Petri net model
– Process Algebras• [Foster, Uchitel, Magee, Kramer ASE’03] Using MSC to model
BPEL web services which are translated to labeled transition systems and verified using model checking
– Model Checking Tools• [Nakajima ICWE’04] Model checking Web Service Flow
Language specifications using SPIN– …
• See the survey on BPEL verification– [Van Breugel, Koshkina 06] Models and Verification of BPEL
http://www.cse.yorku.ca/~franck/research/drafts/
Related Work
• Modeling Choreography & Orchestration
– Process algebras, synchronous communication• [Busi, Gorrieri, Guidi, Lucchi, Zavattaro ICSOC’05]• [Qiu, Zhao, Chao, Yang WWW’07]
– Activity based (rather than message based) approaches• [Berardi, Calvanese, DeGiacomo, Hull, Mecella VLDB’05]