Deterministic simulation testing

Deterministic Simulation
Testing of Distributed Systems
For Great Justice
will.wilson@foundationdb.com
@FoundationDB

2
Debugging distributed
systems is terrible

3
First of all, distributed systems
tend to be complicated…
… but it’s worse than that

4
Repeatability? Yeah, right
A

4
A
A
retry

4
A
A
retry
A

5
Lack of repeatability is a special
case of non-determinism

5
A

5
A
B

5
A A
B

6
FoundationDB is a distributed
stateful system with ridiculously
strong guarantees
ACID!

7
Don’t debug your system,
debug a simulation instead.

8
3 INGREDIENTS
of deterministic simulation
Single-threaded pseudo-concurrency
Simulated implementation of all
external communication
Determinism
1
2
3

Single-threaded 1 pseudo-concurrency
9
Flow
!
A syntactic extension to C++ , new keywords
and actors, “compiled” down to regular C++ w/
callbacks

10
ACTOR Future<float> asyncAdd(Future<float> f, !
! ! ! ! ! ! ! ! ! ! ! float offset) {!
float value = wait( f );!
return value + offset;!
}

11
class AsyncAddActor : public Actor, public FastAllocated<AsyncAddActor> {
public:
AsyncAddActor(Future<float> const& f,float const& offset) : Actor("AsyncAddActor"),
f(f),
offset(offset),
ChooseGroupbody1W1_1(this, &ChooseGroupbody1W1)
{ }
Future<float> a_start() {
actorReturnValue.setActorToCancel(this);
auto f = actorReturnValue.getFuture();
a_body1();
return f;
}
private: // Implementation
…
};
!
Future<float> asyncAdd( Future<float> const& f, float const& offset ) {
return (new AsyncAddActor(f, offset))->a_start();
}

12
int a_body1(int loopDepth=0) {
try {
ChooseGroupbody1W1_1.setFuture(f);
loopDepth = a_body1alt1(loopDepth);
}
catch (Error& error) {
loopDepth = a_body1Catch1(error, loopDepth);
} catch (...) {
loopDepth = a_body1Catch1(unknown_error(), loopDepth);
}
return loopDepth;
}
int a_body1cont1(float const& value,int loopDepth) {
actorReturn(actorReturnValue, value + offset);
return 0;
}
int a_body1when1(float const& value,int loopDepth) {
loopDepth = a_body1cont1(value, loopDepth);
return loopDepth;
}

13
template <class T, class V>
void actorReturn(Promise<T>& actorReturnValue, V const& value) {
T rval = value; // explicit copy, needed in the case the return is a state variable of the actor
Promise<T> rv = std::move( actorReturnValue );
!
// Break the ownership cycle this,this->actorReturnValue,this->actorReturnValue.sav,this-
>actorReturnValue.sav->actorToCancel
rv.clearActorToCancel();
try {
delete this;
} catch (...) {
TraceEvent(SevError, "ActorDestructorError");
}
rv.send( rval );
}

14
Programming language
Ruby (using threads)
Ruby (using queues)
Objective C (using threads)
Java (using threads)
Stackless Python
Erlang
Go
Flow
Time in seconds
1990 sec
360 sec
26 sec
12 sec
1.68 sec
1.09 sec
0.87 sec
0.075 sec

15
Simulated 2 Implementation
INetwork
IConnection
IAsyncFile
…
SimNetwork
SimConnection
SimFile
…

16
void connect( Reference<Sim2Conn> peer, NetworkAddress peerEndpoint ) {
this->peer = peer;
this->peerProcess = peer->process;
this->peerId = peer->dbgid;
this->peerEndpoint = peerEndpoint;
!
// Every one-way connection gets a random permanent latency and a random send buffer
for the duration of the connection
auto latency = g_clogging.setPairLatencyIfNotSet( peerProcess->address.ip, process-
>address.ip, FLOW_KNOBS->MAX_CLOGGING_LATENCY*g_random->random01() );
sendBufSize = std::max<double>( g_random->randomInt(0, 5000000), 25e6 * (latency + .002) );
TraceEvent("Sim2Connection").detail("SendBufSize", sendBufSize).detail("Latency", latency);
}

17
ACTOR static Future<Void> receiver( Sim2Conn* self ) {
loop {
if (self->sentBytes.get() != self->receivedBytes.get())
Void _ = wait( g_simulator.onProcess( self->peerProcess ) );
while ( self->sentBytes.get() == self->receivedBytes.get() )
Void _ = wait( self->sentBytes.onChange() );
ASSERT( g_simulator.getCurrentProcess() == self->peerProcess );
state int64_t pos = g_random->random01() < .5 ? self->sentBytes.get() : g_random-
>randomInt64( self->receivedBytes.get(), self->sentBytes.get()+1 );
Void _ = wait( delay( g_clogging.getSendDelay( self->process->address, self->peerProcess-
>address ) ) );
Void _ = wait( g_simulator.onProcess( self->process ) );
ASSERT( g_simulator.getCurrentProcess() == self->process );
Void _ = wait( delay( g_clogging.getRecvDelay( self->process->address, self->peerProcess-
>address ) ) );
self->receivedBytes.set( pos );
Void _ = wait( Void() ); // Prior notification can delete self and cancel this actor
}
}

18
Simulated Implementation
// Reads as many bytes as possible from the read buffer into [begin,end) and returns the number of
bytes read (might be 0)
// (or may throw an error if the connection dies)
virtual int read( uint8_t* begin, uint8_t* end ) {
rollRandomClose();
!
int64_t avail = receivedBytes.get() - readBytes.get(); // SOMEDAY: random?
int toRead = std::min<int64_t>( end-begin, avail );
ASSERT( toRead >= 0 && toRead <= recvBuf.size() && toRead <= end-begin );
for(int i=0; i<toRead; i++)
begin[i] = recvBuf[i];
recvBuf.erase( recvBuf.begin(), recvBuf.begin() + toRead );
readBytes.set( readBytes.get() + toRead );
return toRead;
}
2

19
Simulated Implementation
void rollRandomClose() {
if (g_simulator.enableConnectionFailures && g_random->random01() < .00001) {
double a = g_random->random01(), b = g_random->random01();
TEST(true); // Simulated connection failure
TraceEvent("ConnectionFailure", dbgid).detail("MyAddr", process-
>address).detail("PeerAddr", peerProcess->address).detail("SendClosed", a > .33).detail("RecvClosed",
a < .66).detail("Explicit", b < .3);
if (a < .66 && peer) peer->closeInternal();
if (a > .33) closeInternal();
// At the moment, we occasionally notice the connection failed immediately. In
principle, this could happen but only after a delay.
if (b < .3)
throw connection_failed();
}
}
2

20
3 Determinism
We have (1) and (2), so all we need to add is
none of the control flow of anything in your
program depending on anything outside of
your program + its inputs.

21
3 Determinism
Use unseeds to figure out whether what
happened was deterministic

23
Test files
testTitle=SwizzledCycleTest
testName=Cycle
transactionsPerSecond=1000.0
testDuration=30.0
expectedRate=0.01
!
testName=RandomClogging
testDuration=30.0
swizzle = 1
!
testName=Attrition
machinesToKill=10
machinesToLeave=3
reboot=true
testDuration=30.0
!
testName=ChangeConfig
maxDelayBeforeChange=30.0
coordinators=auto

24
Other Simulated disasters
In parallel with the workloads, we run
some other things like:
!
• broken machine
• clogging
• swizzling
• nukes
• dumb sysadmin
• etc.

25
You are looking for bugs…
the universe is also
looking for bugs
How can you make your
CPUs more efficient than
the universe’s?

26
1. Disasters here happen
more frequently than in
the real world

27
2. Buggify!
if (BUGGIFY) toSend = std::min(toSend, g_random->randomInt(0, 1000));
!
when( Void _ = wait( BUGGIFY ? Never() : delay( g_network->isSimulated() ? 20 :
900 ) ) ) {
req.reply.sendError( timed_out() );
}

31
We estimate we’ve
run the equivalent of
trillions of hours
of “real world” tests

32
BAD NEWS:
!
We broke debugging (but at least it’s
possible)

34
Two “bugs” found by Sinkhole
1. Power safety bug in ZK
2. Linux package managers writing
conf files don’t sync

35
Future directions
1. Dedicated “red team”
2. More hardware
3. Try to catch bugs that “evolve”
4. More real world testing

Questions?
will.wilson@foundationdb.com
@willwilsonFDB

Data Partitioning
Keyspace is partitioned onto different
machines for scalability

Data Replication
The same keys are copied onto different
machines for fault tolerance

FoundationDB
Partitioning + replication for both scalability and fault
tolerance

Transaction Processing
• Decompose the transaction processing pipeline
into stages
• Each stage is distributed

Pipeline Stages
• Accept client transactions
• Apply conflict checking
• Write to transaction logs
• Update persistent data structures
High Performance!

Performance Results
• 24 machine cluster
• 100% cross-node transactions
• Saturates its SSDs
890,000 ops/sec

Scalability Results
Performance by cluster size

Deterministic simulation testing

Recommended

More Related Content

What's hot (20)

Viewers also liked (13)

Similar to Deterministic simulation testing (20)

Recently uploaded (20)

Deterministic simulation testing