Overview

FIX Antenna C++ performance was measured using the sample from the FA package (samples\Benchmark\Latency2). Configuration properties can also be found in the package.

You can reproduce the measurement on your hardware by compiling and executing the sample described above using the following instruction:

How to measure FA performance using the FA package sample

The following instruction provides the steps to measure the FIX Antenna C++ performance using the Benchmark samples on the FA package.

Two Linux machines should be used for the test, one as the Sender Host (server), the other as the Receiver Host (client).


Download the FIX Antenna C++ package, unzip it and go to the \Samples\Benchmark\Latency2 folder.

  1. Run the makefile.ind to compile all samples.

  2. Go to the \Samples\Benchmark\Latency2\bin folder.
    Open the settings.sh script and change the IP address from the localhost to the IP of the Server Host in the ServerAdress parameter.

    On the Server side run the run.Server.<configuration_name>.sh script and on the Client side - the run.Client.<configuration_name>.sh script. As soon as both of them are executed, the RTT is automatically measured. The latency.csv file with results will appear in the same folder.

    In the optimized.hard configuration the number of cores used by the FIX sessions is directly written in the configuration. Please change it according to your machines. It should be changed in the settings.sh file both in the TasksetServerCores and TasksetClientCores (please see the picture above) parameters. 

    In the optimized.balanced.onload configuration the Solarflare OpenOnload profile is specified. Please specify your existing configuration or set up the new one in the OpenOnload. It should be changed in the settings.sh file both in the OnloadServerProfile and OnloadClientProfile (please see the picture above) parameters. 

Environment

Sender Host:

Receiver Host:

Test scenario

The test scenario is the following:

During the test Round-trip time (RTT) is measured. The first measurement, t1, is made before the message is sent by the Initiator, the second, t2, is made after the received message is parsed by the Initiator.

RTT=t2-t1.

Test configurations

Properties
DefaultAggressiveBalancedOptimized.hard.spinBalanced.onloadOptimized.hard
Prepared messages1
NoNoNoYesNoYes
Affinity
NoNoYesNoYesYes
Thread Pool or Aggressive Mode
Thread Pool
(10 Threads)		Aggressive Mode2Aggressive ModeAggressive ModeAggressive ModeAggressive Mode
Message validation parametersIgnoreUnknownFields





ProhibitUnknownTags✔✔✔✔✔✔
CheckDuplicateTags✔✔✔✔✔✔
MessageMustBeValidated✔✔



VerifyReperatingGroupBounds✔✔



Validation.CheckRequiredGroupFields✔✔



Validation.VerifyTagsValues✔✔



Validation.ProhibitTagsWithoutValue✔✔



Storage type
PersistentMM3PersistentMMPersistentMMPersistentMMPersistentMMPersistentMM
Solarflare OpenOnload4
NoNoNoNoYesYes

1 Prepared messages - a mechanism that preallocates storage for messages (buffers) and then reuses it.

2 Aggressive Mode - 2 dedicated threads are allocated for the FIX session: one for sending messages, another - for receiving them.

3 PersistentMM storage type = a file on the disc.

4  Solarflare OpenOnload - the kernel bypass technique by Solarflare that is activated within the test.


The following description will help to choose the most relevant configuration of FIX Antenna C++:

Results

FIX Antenna C++ configurations comparison

Configuration

DefaultAggressiveBalancedOptimized.hard.spinBalanced.onloadOptimized.hard
RTT (microseconds)

Min

47,146,127,319,011,38,3
Max190,7218,0217,5223,7199,513,2
Average51,651,131,521,613,28,7
RTT distribution (percentiles)
50%49,848,829,921,011,88,5
95%59,959,038,924,018,69,3
99%67,190,445,825,324,610,4

FIX Antenna C++ 2.19.0 vs QuickFIX C++ 1.14.3

The benchmark code was ported to QuickFIX:

struct ExperimentTimes
{
    System::u64 putTime_;
    System::u64 processTime_;
};
class Application :
      public FIX::Application,
      public FIX::MessageCracker
{
public:
 void run() {
  times_ = new ExperimentTimes[10000];
  memset( times_, 0, count*sizeof( ExperimentTimes ) );
  // Get a frequency of timer used
  this->timer_frequency = static_cast<double>(HRTimer::frequency());
  
  FIX42::NewOrderSingle newOrderSingle(
   FIX::ClOrdID( std::to_string(static_cast<long long>(0)) ), FIX::HandlInst( '1' ), FIX::Symbol( "EU" ), FIX::Side( FIX::Side_BUY ),
   FIX::TransactTime(), FIX::OrdType( FIX::OrdType_LIMIT )
  );
  loggedInSem_.wait();
  this->start_time = System::HRTimer::value();
  for( size_t i = 0; i < 10000; ++i ) {
   // Saving send starting time
   times_[putTimeIndex_].putTime_ = HRTimer::value();
   if(putTimeIndex_ < 10000) {
    ++putTimeIndex_;
   }
   newOrderSingle.set( FIX::ClOrdID( std::to_string(static_cast<long long>(i)) ) );
   newOrderSingle.set( FIX::OrderQty( 1 + i%10 ) );
   newOrderSingle.set( FIX::TimeInForce( FIX::TimeInForce_DAY ) );
   newOrderSingle.set( FIX::Price( 30 ) );
   newOrderSingle.getHeader().setField( FIX::SenderCompID( "INCA" ) );
   newOrderSingle.getHeader().setField( FIX::TargetCompID( "TW" ) );
   FIX::Session::sendToTarget(newOrderSingle);
   // 1K messages per second
   double next_send_time = static_cast<double>(i+1)/1000;
   while(true){
    //Wait for next sending time
    double current_time = static_cast<double>(System::HRTimer::value() - start_time)/timer_frequency;
    double to_sleep = next_send_time - current_time;
    if(to_sleep < 0.0)
     break;
    if(0.025 < to_sleep)
                    Thread::sleep( static_cast<int>(1000.0*0.9*to_sleep) );
   }
  }
 }
private:
 ...
 ExperimentTimes* times_;
 size_t putTimeIndex_;
 size_t processTimeIndex_;
 
 double timer_frequency;
 
 Semaphore loggedInSem_;
 ...
 
 void onLogon( const FIX::SessionID& sessionID ) {
  loggedInSem_.post();
 }
 
 void fromApp( const FIX::Message& message, const FIX::SessionID& sessionID )
      throw( FIX::FieldNotFound, FIX::IncorrectDataFormat, FIX::IncorrectTagValue, FIX::UnsupportedMessageType ) {
          crack( message, sessionID );
 };
 void onMessage( const FIX42::ExecutionReport&, const FIX::SessionID& ) {
  // Saving message receiving time
  times_[processTimeIndex_].processTime_ = HRTimer::value();
        processTimeIndex_++;
 };
 ...
}
class Application :
      public FIX::Application,
      public FIX::MessageCracker
{
 FIX42::ExecutionReport answer;
 int uniqueId;
 ...
 void onMessage( const FIX42::NewOrderSingle&, const FIX::SessionID& ) {
  FIX::SenderCompID senderCompID;
  FIX::TargetCompID targetCompID;
  FIX::ClOrdID clOrdID;
  FIX::Symbol symbol;
  FIX::Side side;
  FIX::OrdType ordType;
  FIX::Price price;
  FIX::OrderQty orderQty;
  FIX::TimeInForce timeInForce( FIX::TimeInForce_DAY );
  message.getHeader().get( senderCompID );
  message.getHeader().get( targetCompID );
  message.get( clOrdID );
  message.get( symbol );
  message.get( side );
  message.get( ordType );
  if ( ordType == FIX::OrdType_LIMIT )
   message.get( price );
  message.get( orderQty );
  message.getFieldIfSet( timeInForce );
  answer.set( clOrdID );
  answer.set( price );
  answer.set( orderQty );
  answer.set( FIX::TransactTime() );
  answer.set( FIX::ExecID(std::to_string(static_cast<long long>(uniqueId)) ) );
  answer.set( FIX::OrderID(std::to_string(static_cast<long long>(uniqueId)) ) );
  uniqueId++;
  answer.set( FIX::LeavesQty(orderQty) );
  FIX::Session::sendToTarget(answer);
 };
 ...
}


Results can be compared in the following table:

Configuration

FIXAntenna defaultFIXAntenna balancedQuickFIX no validationQuickFIX validation
RTT (microseconds)

Min

47,127,3155,0157
Max190,1217,5371,5377,8
Average51,631,5158,9161,8
RTT distribution (Percentiles)
50%49,829,9157,6159,5
95%59,938,9165,0167,1
99%67,145,9177,4183,5