EPOBF: Energy Efficient Allocation of Virtual Machines in High Performance Computing Cloud - Transactions on Large-Scale Data-and Knowledge-Centered Systems - page 82

Fig. 3. Length in Millions Instructions (MI) of the 48,880 simulated cloudlet. The
y-axis is the length of cloudlets (in x10000 Million Instructions) in log-scaled values.
5.1 Results and Discussions
We evaluate these heuristics by simulations using a simulated cloud data cen-
ter system that has 10,000 heterogeneous hosts (there are three groups of hosts
with different cores and physical memory), and some simulated HPC workload
with thousands of CloudSim's cloudlets [ 9 ] (we assume that each HPC job's
task is modeled as a cloudlet). The information of virtual machines (and also
cloudlets) in these simulated workloads is extracted from a real log-trace (SDSC-
BLUE-2000-4.1-cln.swf [ 3 ]) in Feitelson's Parallel Workloads Archive (PWA) [ 2 ]
to model HPC jobs (original job's executing time adds 300 seconds that is over-
head of using each VM). When converting from the log-trace, each cloudlet's
length is a product of the system's processing time and CPU rating (we set the
CPU rating is 375). We assign job's submission time, start time, and execution
time, and number of processors in job data from the SDSC-BLUE-2000-4.1-cln
to cloudlet's submission time, starting time and cloudlet's length, and num-
ber of cloudlets. We use four types of VMs (e.g. Amazon EC2's VM instances:
high-CPU, extra, small, micro): each VM has only one core and maximum per-
formance of VM is 2500, 2000, 1000, 500 MIPS, 870, 3840, 1740, 613 MB of
RAM and network of 10000 Kbits/s. We choose the latest version (version
3.0) of the CloudSim [ 9 ] to model and simulate our HPC cloud and the VM
allocation heuristics. The CloudSim's framework is used to write and evalu-
ate new customized allocation algorithms for VMs and cloudlets. We based on
the previous work [ 6 ] to improve existing PABFD allocation heuristic for this
paper.
We show here the results of simulations using a simulated HPC workload with
48,880 cloudlets, in which parameters of all cloudlets are taken from 800 jobs in