Moldable Job Allocation for Handling Resource Fragmentation in Computational Grid - Cloud, Grid and High Performance Computing: Emerging Applications - page 64

Table 1. Characteristics of the workload log on SDSC's SP2

Number of
jobs

Maximum
execution time
(sec.)

Average
execution time
(sec.)

Maximum number of
processors
per job

Average number of processors
per job

Queue 1 4053 21922 267.13 8 3

Queue 2 6795 64411 6746.27 128 16

Queue 3 26067 118561 5657.81 128 12

Queue 4 19398 64817 5935.92 128 6

Queue 5 177 42262 462.46 50 4

Total 56490

• Waiting time . It is the time between a
job's submission and its allocation.
• Actual runtime . When moldable job al-
location is applied, a job's actual runtime
may be different from the runtime recorded
in the workload log. This attribute records
the actual runtime it takes.
• Actual number of processors . When the
scheduler applies moldable job allocation,
the number of processors a job actually
uses for execution may be different from
the value recorded in the workload log.
This attribute records the number of pro-
cessors actually used.
• Execution site . In a computational Grid
environment, a job may be scheduled to
run on a site other than its home site. The
attribute records the actual site that it runs
on.
• Turnaround time . The simulation pro-
gram calculates each job's turnaround time
after its execution and records the value in
this attribute.

Our simulation studies were based on publicly
downloadable workload traces (“Parallel Work-
loads Archive,”). We used the SDSC's SP2 work-
load logs 1 on (“Parallel Workloads Archive,”) as
the input workload in the simulations. The detailed
workload characteristics are shown in Table 1.
In the SDSC's SP2 system the jobs in the logs
are put into different queues and all these queues
share the same 128 processors. In section 4, this
original workload is directly used to simulate a
homogeneous parallel computer with 128 proces-
sors. In section 5 the workload log will be used
to model the workload on a computational Grid
consisting of several different sites whose work-
loads correspond to the jobs submitted to the
different queues respectively. Table 2 shows the
configuration of the computational Grid accord-
ing to the SDSC's SP2 workload log. The number
of processors on each site is determined according
to the maximum number of required processors
of the jobs belonged to the corresponding queue
for that site.
To simulate the speed difference among par-
ticipating sites we define a speed vector, e.g.
speed=(sp1,sp2,sp3,sp4,sp5), to describe the

Table 2. Configuration of the computational Grid according to SDSC's SP2 workload

total site 1 site 2 site 3 site 4 site 5

Number of processors 442 8 128 128 128 50