Information Technology Reference
In-Depth Information
Algorithm 3.
universe = vanilla
executable = runjob
output = stdout.log
error = stderr.log
log = job.log
transfer_input_files = <executable>,descrgen.so
...
all CEs, it has to be sneaked in the target machine
along with the job. The Condor classAD is pre-
pared in this respect so that a
simple shell
script
setting the LD_PREALOAD environment vari-
able and running the job will be executed by the
work node. The job and the shared object of the
Description generator are transferred as input files
of the job runner executable. The classAD frag-
ment in Algorithm 3 demonstrates the technique.
shortest path searching algorithm in a graph. The
job first parses the graph description loaded from
a 300 MB input file then it starts to calculate the
shortest path from the given parameter node to
every other node in the graph. The jobs provided
by OptorSim are using input files of 10 GB each.
The number of jobs was also raised to 500 and
1000, to provide us with sufficient job queue sizes
on the CEs. Before the simulation was performed
OptorSim was supplied with the necessary com-
plex job descriptions.
Due to the lack of support for querying some
file related information from the current P-Grade
portal, we have simulated mainly solutions that
do not use such information during the schedul-
ing process (the default schedulers in this scope
in OptorSim are the Random and Queue Length
strategies). The static data feeder strategy has been
simulated both using and not using file informa-
tion, while the dynamic data feeder strategy was
simulated only with file information present. The
benefits can be clearly seen. The static data feeder
algorithm performs significantly better when the
correct size of the files used by the jobs is known
(Static DF) compared to the scheduling when
information about expected file transfer times is
absent (Static DF no FS info).
According to the simulation results (see Figure
3 for the mean job completion time values pro-
vided by OptorSim), using the static data feeder
scheduler (Static DF no FS info) the mean job
completion time of all jobs on Grid is about 3-4%
lower than in the case of the schedulers which do
Description Repository Service
For similar reasons which do not allow the perma-
nent deployment of the
Description generator
, the
Description repository service
cannot be deployed
inside the Grid either. Therefore, we have not
implemented it in the current Grid environment
supported by P-GRADE.
SIMULATION RESULTS
Simulations were conducted by using OptorSim
v2.0 (Simulating data access optimization algo-
rithms - OptorSim), which was extended with
the proposed static data feeder and dynamic data
feeder scheduler implementations. The extended
OptorSim was configured to use the EDG topology
specified by the configuration file shipped with
the simulator. The CEs of the configuration were
extended with MIPS values. One of the group of
jobs submitted to the Grid (approx. 1/3 of the total
jobs) was changed to simulate the single source
Search WWH ::
Custom Search