![](bg5.png)
below.
1. While a job is being scheduled by SSS, SSS
places a data request to Local Data Manager,
which is the LDMS’s service point to SSS.
LDM puts this request in a queue and invokes
LDMS.
2. For the current data request, LDMS contacts
with Local Replica Location Service and learns
if the requested data item(s) can be locally
provided.
3. For the data items which are already available
in the site, LDMS submits a reservation request
to Local Reservation Service. This request
includes the all the links from a local storage
element to the chosen computing element,
reservation start (current time) and finish
(request deadline) times, and bandwidth value
(bw).
4. For the unfound data items, LDMS sends a
data request to Data Manager so that these
items can be copied from some remote site(s)
into this site.
5. If all local reservations are successfully made,
and/or LDMS is informed by Data Manager
that the requested data item(s) will be available
at the task start time, Local Data Manager
notifies SSS with either positive/negative
acknowledgement accordingly.
5 EXPERIMENTS
Using DGridSim, the three Grid scheduling
algorithms were evaluated. With the start of the
simulation, a Data Grid system was created. The
system was assumed to have the following
properties. It has ten sites each of which includes
thirty-two heterogeneous computing elements and a
single storage element. The computing elements
have MIPS rating of U~[7500, 12500], where U~
means uniformly distributed, and the storage
elements have storage capacity of U~[175000,
225000] Mbytes. Furthermore, every site is
equipped with a gateway router to which all
computing and storage elements are connected. The
links between computing and storage elements and
their gateway have bandwidth of U~[750, 1250]
Mbytes/sec and U~[1750, 2250] Mbytes/sec,
respectively. Ten gateway routers are interconnected
by a randomly generated network topology
composed of ten routers and twenty-five links whose
bandwidths with an average bandwidth of U~[750,
1250] Mbit/sec.
After the creation of a Data Grid system, jobs
were produced. The jobs were characterized with job
size, deadline, and the number of data items. In the
base set of simulation studies, job sizes are
U~[3750000, 7500000] MI (Million Instruction),
deadlines are U~[750, 1250] seconds, and the
number of data items needed in order for jobs to
start their execution is just one. During the
simulations, jobs were submitted to the system with
a rate of one job per five seconds.
Initially, all two-hundred different data items
were assumed to be stored in a single (Tier-0) site
without any computing elements. Thus, all data are
distributed from this Tier-0 site to all other sites with
computing capability. Moreover, jobs are randomly
associated with the data items whose sizes are
U~[750, 1250] Mbytes.
Using DGridSim, a base set of results was first
established for the following parameter values:
Number of Jobs=1000, Mean Job Size=5000000 MI,
Mean Job Deadline=100 sec, Mean Number of Data
Items=1 and Mean Link Bandwidth=125
Mbytes/sec. Later, these parameters are varied and
the effects are observed and reported in Tables 1-3.
Each data in Tables 1-3 denotes the average
satisfiability (the ratio of number of jobs finished
before their deadlines to total number of jobs) in
three simulation runs.
Table 1 shows the effect of changing the number
of jobs submitted to the system from 1000 to 3000
on the algorithms. According to Table 1, all three
algorithms maintain a relatively stable performance
on average. Furthermore, Random and EDF have
shown very similar performance, and they are
slightly better than MCTF.
Table 1: Performance of the three grid scheduling
algorithms when the number of jobs is increased.
Number of Jobs
1000 1500 2000 2500 3000
Random 0.94 0.96 0.96 0.97 0.98
EDF 0.94 0.96 0.96 0.97 0.98
MCTF 0.92 0.90 0.89 0.92 0.85
Table 2 shows the impact of increasing the mean
job size from 5000000 (5 M) MI to 1000000 (10 M)
MI on the algorithms. According to Table 2,
increasing job sizes significantly degrades the
performance of MCTF. On the other hand, it seems
that both Random and EDF keep its performance at
a top level.
SIMULTECH 2011 - 1st International Conference on Simulation and Modeling Methodologies, Technologies and
Applications
178