Adaptive Execution of Scientific Workflow Applications on Clouds - Cloud Computing with e-Science Applications - page 85

However, we noticed that the time-based simulation has limited scalabil-
ity. When executed in a private cloud of 50 VMs, it took 35,248 seconds to
complete the 2 million commuters' agent-based simulation. This happened
because there were dependencies in t + 1 with time t (i.e., simulation at time
t needs to be completed before simulation of time t + 1 starts). Because of this
issue, we replaced the time-based simulation with an event-based simulation.
In event-based simulation, the model handles the agents' interactions, such
as boarding of commuters, unboarding of commuters, train arrivals at sta-
tions, and train departures from stations. On the back end, the workload is
distributed via a similar method to other phases (each process handles a
group of agents). With this new technique, the execution time of the simula-
tion in the same private cloud was completed in 1,818 seconds for the same
2-million-commuter agent-based simulation, an improvement of 19 times
over the original technique.
We further scaled the agent-based simulation by executing it on 1,000 VMs.
In this case, the agent-based simulation completed in 434 seconds for simula-
tion of 2 million commuters and 963 seconds for 7 million commuters. This
demonstrated that the three phases of our approach are scalable and suitable
for execution on elastic cloud platforms.
To summarize, we gave preference to the cloud-enabled WfMS over the
ZMQ system because of the following reasons: (1) It enabled more efficient
management of the highly distributed data required by the agent-based
simulation workflow; (2) it better automated the workflow process for data
analytics with multiobjective optimization of performance and budget;
and (3) it enabled dynamic resource allocation for adaptive services with
fault tolerance.
4.7 Related Work
Given the importance of workflow applications for the scientific community,
many scientific workflow platforms were developed to explore scientific
computational platforms such as grids. As cloud platforms became popular
among the scientific community, WfMSs where enhanced to support them.
Pegasus [1] offers a set of tools for different aspects of execution and man-
agement of workflow applications and platforms. It implements application
programming interfaces (APIs) for diverse programming languages, supports
submission of workflows via web portals, and integrates with external tools. On
its back end, it supports multiple cloud providers and scientific infrastructures.
Taverna [2] is another widely adopted workflow engine that can explore
both grid and cloud platforms. Applications running on the platform can
be deployed in many modes, including “server mode,” by which it supports
requests from many users to execute remote workflow applications.