[omniORB] Strange performance result

[Zhao, Jason]

Hi,

I'm sorry if this problem looks complicated.

I used a simple idl to test omniorb's performance on oneway calls to
pass data between two machines. When the bandwidth between the two
machines is less than 150Mbps, the performance results are consistent:
different runs produced similar numbers and within in the same run,
larger message sizes result in higher application level throughput. But
when the bandwidth between the two machines are set to be higher than
150Mbps (using Linux traffic control and physical interface is gig
Ethernet), the result numbers are unpredictable: not consistent across
multiple runs and even within the same run, larger message sizes some
times result in lower application level throughput.

I tested ACE TAO using the same idl and under the same environment, the
results are always consistent either across runs or within the same run.
Does anyone know what might caused omni to behave unpredictably when the
bandwidth is higher than 150Mbps (cpu utilization was not very high in
those cases)?

Thank you.

Jason


Configuration:

OMNIORB_4_1_0_BETA_1 snapshot got from
http://omniorb.sourceforge.net/snapshots/omniORB-4.1-latest.tar.gz on
05/15
Ipv6
Linux version 2.4.21-4.ELsmp (bhcompile at daffy.perf.redhat.com) (gcc
version 3.2.3 20030502 (Red Hat Linux 3.2.3-20)) #1 SMP
Fri Oct 3 17:52:56 EDT 2003
CPU * Intel Xeon 2.66Ghz 533Mhz FSB 512K Cache
MEMORY * 2x 512MB DDR266 PC2100 ECC Reg Memory

===============
Here is the idl used (from ACE TAO's performance test suit), I tested
message sizes of 128, 256, 512, 1024, 2048, and 4096 bytes, sent 10000
messages for each size and calculated application level throughput per
message size.

module Test
{
  /// The data payload
  typedef sequence<octet> Payload;
  struct Message {
    unsigned long message_id;
    Payload the_payload;
  };

  /// Implement a simple interface to receive a lot of data
  interface Receiver
  {
    /// Receive a big payload
    oneway void receive_data (in Message the_message);

    /// All the data has been sent, print out performance data
    void done ();
  };

  /// Implement a factory to create Receivers
  interface Receiver_Factory
  {
    /// Create a new receiver
    Receiver create_receiver ();

    /// Shutdown the application
    oneway void shutdown ();
  };
};

================
Here is the omniorb configuration file used (comments removed for
message size issue), basically I make the server side single-threaded so
that messages were processed in the order they were sent.

traceLevel = 0
traceExceptions = 0
traceInvocations = 0
traceInvocationReturns = 0
traceThreadId = 0
traceTime = 0
dumpConfiguration = 0
maxGIOPVersion = 1.2
giopMaxMsgSize = 2097152    # 2 MBytes.
strictIIOP = 0
tcAliasExpand = 0
useTypeCodeIndirections = 1
acceptMisalignedTcIndirections = 0
scanGranularity = 5
nativeCharCodeSet = ISO-8859-1
nativeWCharCodeSet = UTF-16
abortOnInternalError = 0
abortOnNativeException = 0
InitRef = NameService=corbaname::[2001:411:2:6:2::]
DefaultInitRef = corbaloc::[2001:411:2:6:2::]
clientCallTimeOutPeriod = 0
clientConnectTimeOutPeriod = 0
supportPerThreadTimeOut = 0
outConScanPeriod = 120
maxGIOPConnectionPerServer = 1
oneCallPerConnection = 1
offerBiDirectionalGIOP = 0
diiThrowsSysExceptions = 0
verifyObjectExistsAndType = 0
giopTargetAddressMode = 0
bootstrapAgentPort = 900
endPoint = giop:tcp:[2001:411:2:6:2::]:20000
serverCallTimeOutPeriod = 0
inConScanPeriod = 180
threadPerConnectionPolicy = 1
maxServerThreadPerConnection = 1
maxServerThreadPoolSize = 100
threadPerConnectionUpperLimit = 10000
threadPerConnectionLowerLimit = 9000
threadPoolWatchConnection = 1
connectionWatchPeriod = 50000
connectionWatchImmediate = 0
acceptBiDirectionalGIOP = 0
unixTransportDirectory = /tmp/omni-%u
unixTransportPermission = 0777
supportCurrent = 0
copyValuesInLocalCalls = 0
objectTableSize = 100
poaHoldRequestTimeout = 0
poaUniquePersistentSystemIds = 1
supportBootstrapAgent = 0