It seems that when Globus jobs with larger (>240K) executables are
submitted (either via Condor-G or globus-job-run), the jobmanager on the
server stalls on stage_in phase with globus-gass-cache-util running in a
tight loop:

select(10, [3 9], [], NULL, NULL)       = 1 (in [9])
gettimeofday({1096921850, 657981}, NULL) = 0
select(10, [3 9], [], NULL, NULL)       = 1 (in [9])
gettimeofday({1096921850, 658569}, NULL) = 0
select(10, [3 9], [], NULL, NULL)       = 1 (in [9])
gettimeofday({1096921850, 659157}, NULL) = 0
select(10, [3 9], [], NULL, NULL)       = 1 (in [9])
gettimeofday({1096921850, 659744}, NULL) = 0
select(10, [3 9], [], NULL, NULL)       = 1 (in [9])
gettimeofday({1096921850, 660386}, NULL) = 0
select(10, [3 9], [], NULL, NULL)       = 1 (in [9])
gettimeofday({1096921850, 660974}, NULL) = 0
select(10, [3 9], [], NULL, NULL)       = 1 (in [9])
gettimeofday({1096921850, 661562}, NULL) = 0
select(10, [3 9], [], NULL, NULL)       = 1 (in [9])
gettimeofday({1096921850, 662152}, NULL) = 0
select(10, [3 9], [], NULL, NULL)       = 1 (in [9])
gettimeofday({1096921850, 662740}, NULL) = 0

... etc

This works fine with smaller executables.