pip: detecting the unexpected in distributed systems charles killian amin vahdat ucsd patrick...

Pip: Detecting the Unexpected in

Distributed Systems

Charles KillianAmin Vahdat

UCSD

Patrick [email protected]

u

Collaborators:Janet WienerJeffrey MogulMehul Shah

HP Labs

http://issg.cs.duke.edu/pip/

NSDI - May 8th, 2006

Introduction

• Distributed systems are essential to the Internet• Distributed systems are complex

– Harder to debug than centralized systems

• Pip uses causal paths to find bugs– Characterize system behavior– Deviations from expected behavior often indicate bugs

• Experimental results show that Pip is effective– Found 19 bugs in 6 test systems


Challenges of debugging distributed systems

• Distributed systems have more bugs– Parallelism is hard– New sources of failure

• More components = more faults

• Network errors

• Security breaches

• Finding bugs is harder– More nodes, events, and messages to keep track of– Applications may cross administrative domains– Bugs on one node may be caused by events on

another


Causal path analysis

• Programmers wish to examine and check system-wide behaviors– Causal paths– Components of end-to-end

delay– Attribution of resource

consumption• Unexpected behavior might indicate a bug– Pip compares actual behavior

to expected behavior

Web server

App server

Database

500ms

2000pagefaults


What kinds of bugs?

• Structure bugs– Incorrect placement/timing of processing

and communication

• Performance bugs– Throughput bottlenecks– Over- or under-consumption of resources

• Bugs present in an actual run of the system– Selecting input for path coverage is

orthogonal

Web server

App server

Database

500ms

2000pagefaults


Three target audiences

• Primary programmer– Debugging or optimizing his/her own system

• Secondary programmer– Inheriting a project or joining a programming

team– Discovery: learning how the system behaves

• Operator– Monitoring a running system for unexpected

behavior– Performing regression tests after a change


Outline

• Introduction• Pip

– Expressing expected behavior– Exploring application behavior– Results

• Conclusions


Workflow

Pip:1. Captures events from a running

system2. Reconstructs behavior from events3. Checks behavior against

expectations4. Displays unexpected behavior

Both structure and resource violations

Goal: help programmers locate and explain bugs

Behaviormodel

Application

Application

Expectations

Pip checker

Unexpectedstructure

Resourceviolations

Pip explorer: visualization GUI


Describing application behavior

• Application behavior consists of paths– All events, on any node, related to one high-level

operation– Definition of a path is programmer defined– Path is often causal, related to a user request

WWW

App srv

DB

Parse HTTP

Query

Send response

Run application

time


Describing application behavior

• Within paths are tasks, messages, and notices– Tasks: processing with start and end points– Messages: send and receive events for any

communication• Includes network, synchronization (lock/unlock), and

timers

– Notices: time-stamped strings; essentially log entries

WWW

App srv

DB

Parse HTTP

Query

Send response

Run application

time

“Request = /cgi/…” “2096 bytes in response”

“done with request 12”


Outline

• Introduction• Pip

– Expressing expected behavior– Exploring application behavior– Results

• Conclusions


Expectations

• Pip has two kinds of expectations

• Recognizers classify paths into sets

• Aggregates assert properties of sets

R1

R2

R3

?


Expectations: recognizers

• Application behavior consists of paths• Each recognizer matches paths

– A path can match more than one recognizer• A recognizer can be a validator, an invalidator, or

neither• Any path matching zero validators or at least one

invalidator is unexpected behavior: bug?validator CGIRequest thread WebServer(*, 1) task(“Parse HTTP”) limit(CPU_TIME, 100ms); notice(m/Request URL: .*/); send(AppServer); recv(AppServer);

invalidator DatabaseError notice(m/Database error: .*/);


Expectations: recognizers language

• repeat: matches a ≤ n ≤ b copies of a block

• xor: matches any one of several blocks

• future: lets a block match now or later– done: forces the named block to match

repeat between 1 and 3 { … }

xor {branch: …branch: …

}

future F1 { … }…done(F1);


Expectations: aggregate expectations

• Recognizers categorize paths into sets• Aggregates make assertions about sets of paths

– Instances, unique instances, resource constraints– Simple math and set operators

assert(instances(CGIRequest) > 4);assert(max(CPU_TIME, CGIRequest) < 500ms);assert(max(REAL_TIME, CGIRequest) <= 3*avg(REAL_TIME, CGIRequest));


Automating expectations and annotations

• Expectations can be generated from behavior model– Create a recognizer for each

actual path– Eliminate repetition– Strike a balance between

over- and under-specification

• Annotations can be generated by middleware– As in several of our test

systems

Behaviormodel

Application

Application

Expectations

Pip checker

Annotations

Unexpectedbehavior


Exploring behavior

•Expectations checker generates lists of valid and invalid paths

•Explore both sets– Why did invalid paths occur?– Is any unexpected behavior misclassified as

valid?• Insufficiently constrained expectations•Pip may be unable to express all expectations


Timing and resource properties for one taskCausal view of path

Visualization: causal paths

Caused tasks, messages, and notices on that thread


Visualization: timelines

• View one path instance as a timeline– Different colors = different hosts– Each bar = task – Click a bar to see task details


Visualization: performance graphs

• Plot per-task or per-path resource metrics– Cumulative distribution (CDF)– Probability density (PDF)– Changing values over time

• Click on a point to see its value and the task/path represented

Time (s)

Dela

y (

ms)


Results

• We have applied Pip to several distributed systems:– FAB: distributed block store– SplitStream: DHT-based multicast protocol– Others: RanSub, Bullet, SWORD, Oracle of Bacon

• We have found unexpected behavior in each system

• We have fixed bugs in some systems… and used Pip to verify that the behavior was fixed


Results: SplitStreamDHT-based multicast protocol

13 bugs found, 12 fixed– 11 found using expectations, 2 found using GUI

• Structural bug: some nodes have up to 25 children when they should have at most 18– This bug was fixed and later reoccurred– Root cause #1: variable shadowing– Root cause #2: failed to register a callback

• How discovered: first in the explorer GUI, confirmed with automated checking


Results: FABDistributed block store

1 bug found, fixed– Four protocols checked: read, write, Paxos,

membership

• Performance bug: quorum operations call nodes in arbitrary order– Should overlap computation when possible


Results: FABDistributed block store

• For disk-bound workloads, call self last• For cache-bound workloads, call self second

– Actually, last in quorum


Conclusions

• Causal paths are a useful abstraction of distributed system behavior

• Expectations serve as a high-level description– Summary of inter-component behavior and timing– Regression test for structure and performance

• Finding unexpected behavior can help us find bugs– Both structure and performance bugs– Visualization can expose additional bugs

http://issg.cs.duke.edu/pip/

NSDI - May 8th, 2006 26

Extra slides

• Related work• Pip results: RanSub• Pip vs. printf• Pip resource metrics• Building a behavior model• Annotations• Checking expectations• Visualization: communication graph


Related work

• Causal paths– Pinpoint [Chen, 2004]– Magpie [Barham, 2004]

• Black-box inferences– Finding stepping stones [Zhang, 1997]– Wavelets to debug network performance [Huang,

2001]• Expectations-checking systems

– PSpec [Perl, 1993]– Meta-level compilation [Engler, 2000]– Paradyn [Miller, 1995]

• Model checking– MaceMC [Killian, 2006]– VeriSoft [Godefroid, 2005]


Pip results: RanSub

2 bugs found, 1 fixed• Structural bug: during first round of communication, parent nodes send summary messages before hearing from all children– Root cause: uninitialized state variables

• Performance bug: linear increase in end-to-end delay for the first ~2 minutes– Suspected root cause: data structure listing all

discovered nodes


Pip vs. printf

• Both record interesting events to check off-line– Pip imposes structure and automates checking– Generalizes ad hoc approaches

Pip printfNesting, causal order UnstructuredTime, path, and thread No contextCPU and I/O data No resource informationAutomatic verification using declarative language

Verification with ad hoc grep or expect scripts

SQL queries “Queries” using Perl scripts

Automatic generation for some middleware

Manual placement


Pip resource metrics

• Real time• User time, system time

– CPU time = user + system– Busy time = CPU time / real time

• Major and minor page faults (paging and allocation)• Voluntary and involuntary context switches• Message size and latency• Number of messages sent• Causal depth of path• Number of threads, hosts in path


Building a behavior model

Sources of events:• Annotations in source code

– Programmer inserts statements manually• Annotations in middleware

– Middleware inserts annotations automatically– Faster and less error-prone

• Passive tracing or interposition– Easier, but less information

• Or any combination of the above

Model consists of paths constructed from events recorded by the running application


Annotations

• Set path ID• Start/end task• Send/receive message• Notice

WWW

App srv

DB

Parse HTTP

Query

Send response

Run application

time

“Request = /cgi/…” “2096 bytes in response”

“done with request 12”


Checking expectations

Traces

Categorized paths

Reconciliation

Events database

Paths

Path construction

Expectation checking

Application

Application

For each path P For each recognizer R Does R match P?Check each aggregate

Expectations

Match start/end task, send/receive message

Organize events into causal paths


Visualization: communication graph

• Graph view of all host-to-host network traffic– Nodes = hosts, edges = network links in use– Directed edge = unidirectional communication

pip: detecting the unexpected in distributed systems charles killian amin vahdat ucsd patrick...

Documents

actual behavior

eventschecks behavior

pathsall events

kinds of bugs

path coverage

pathsa path

internetdistributed

resource violationsgoal