The following experiments were run on the nomer.mcs.anl.gov virtual cluster. The cluster consists of 6 partitions, each having a single Intel(R) Xeon(R) CPU E5430 @ 2.66GHz core, and 2GB RAM. The virtual machines in the cluster each had a single virtual network interface. The cluster was configured as follows:

All nodes ran an apache2 http server, gmond (ganglia monitoring), and pbs_mom (torque LRM).

The master node also ran a globus-gatekeeper, globus-gridftp-server, globus-job-manager-event-generator, gmetad (Ganglia Meta Daemon), pbs_sched (Torque LRM scheduler), pbs_server (Torque LRM server), and nfsd linux kernel NFSv4 server for the execution nodes.

The test nodes (1 for single-client tests, 5 for multiple client tests) ran the gram-throughput-tester program. All 5 test/execution nodes executed the test job executables as scheduled by the LRM.

These tests were done with GT 5.0.0 beta1. The throughput tester was compiled from CVS trunk October 30, 2009.

The first set of tests submitted jobs to the GRAM5 service for one hour. After one hour elapsed, the client would terminate any jobs which were being managed by the GRAM5 service. The test client recorded the time of the experiment start, the time of the termination start, and the time after which all jobs had reached a DONE or FAILED state.

The throughput test client would generate a maximum of 50 simultaneous job requests. For all but the uncapped test below, a maximum of 2000 jobs were managed by the job manager at any given time (pending or active). Once the client had submitted 2000 jobs, it would stop submissions until a job completed. A total of 10 execution slots were available for the LRM to schedule jobs into, so many were in the GRAM5 PENDING state during the duration of the test.

The different gram-throughput-tester experiments consisted of:

The following table contains a summary of the results of these experiments. The columns contain the following information:

The following experiments were run on the nomer.mcs.anl.gov virtual cluster. The cluster consists of 6 partitions, each having a single Intel(R) Xeon(R) CPU E5430 @ 2.66GHz core, and 2GB RAM. The virtual machines in the cluster each had a single virtual network interface. The cluster was configured as follows:

All nodes ran an apache2 http server, gmond (ganglia monitoring), and pbs_mom (torque LRM).

The master node also ran a globus-gatekeeper, globus-gridftp-server, globus-job-manager-event-generator, gmetad (Ganglia Meta Daemon), pbs_sched (Torque LRM scheduler), pbs_server (Torque LRM server), and nfsd linux kernel NFSv4 server for the execution nodes.

The test/execution node ran the condor-g daemons. The condor job classified ad included attributes to submit the jobs to the GRAM5 service running on the service node. The tests were done with Condor version 7.4.1. The Condor-G configuration parameters in the condor_config file were as follows:

GRIDMANAGER_MAX_PENDING_SUBMITS_PER_RESOURCE=50
GRIDMANAGER_MAX_SUBMITTED_JOBS_PER_RESOURCE=2000
GRIDMANAGER_MAX_PENDING_REQUESTS=50
GRIDMANAGER_JOB_PROBE_INTERVAL=300
GRIDMANAGER_MAX_JOBMANAGERS_PER_RESOURCE=0
ENABLE_GRID_MONITOR=FALSE
GRIDMANAGER_DEBUG= D_FULLDEBUG
GRIDMANAGER_GLOBUS_COMMIT_TIMEOUT=12000

The execution nodes executed the test job executables as scheduled by the LRM.

For this test, the test/execution node and the execution nodes where configured to run up 20 job processes each simultaneously.

This test submitted a 2000 job condor job cluster, using the following classified ad:

Universe=grid
grid_resource = gt5 nomer1.mcs.anl.gov:2119/jobmanager-pbs
executable=/bin/sleep
arguments=300
transfer_executable=False
stream_output = False
stream_error  = False
output = test.out.$(Process)
error  = test.err.$(Process)
log    = test.log
notification=Never
queue 2000

The configuration parameters are similar to the GRAM5 tests described in GRAM5 Throughput Tests section. The key difference being that this test runs until all 2000 jobs have completed and does not submit any jobs after the maximum of 2000 has been reached.

To provide a point of comparison, another test using similar parameters was run using the gram-throughput-tester program in place of Condor-G. Note that the Condor-G service provides file staging and a scratch directory beyond what the throughput tester job did.

The two experiments consist of:

The following table contains a summary of the results of these experiments. The columns contain the following information: