Fine-Grained Profiling for Data-Intensive Workflows.
ABSTRACT Profiling is an effective dynamic analysis approach to investigate complex applications. ParaTrac is a user-level profiler using file system and process tracing techniques for data-intensive workflow applications. In two respects ParaTrac helps users refine the orchestration of workflows. First, the profiles of I/O characteristics enable users to quickly identify bottlenecks of underlying I/O subsystems. Second, ParaTrac can exploit fine-grained data-processes interactions in workflow execution to help users understand, characterize, and manage realistic data-intensive workflows. Experiments on thoroughly profiling Montage workflow demonstrate that ParaTrac is scalable to tracing events of thousands of processes and effective in guiding fine-grained workflow scheduling or workflow management systems improvements.
- SourceAvailable from: Dennis Gannon[Show abstract] [Hide abstract]
ABSTRACT: Workflows have been used to model repeatable tasks or operati ons in a number of different industries includ- ing manufacturing and software. In recent years, workflows a re increasingly used in distributed resources and web services environments through resource models such as grid and cloud computing. These workflows often have disparate requirements and constraints that nee d to be accounted for during workflow orchestra- tion. In this paper, we present workflow examples from differ ent domains including bioinformatics and biomedical, weather and ocean modeling, astronomy detailing their data and computational requirements.
Conference Paper: Characterization of scientific workflows[Show abstract] [Hide abstract]
ABSTRACT: Researchers working on the planning, scheduling and execution of scientific workflows need access to a wide variety of scientific workflows to evaluate the performance of their implementations. We describe basic workflow structures that are composed into complex workflows by scientific communities. We provide a characterization of workflows from five diverse scientific applications, describing their composition and data and computational requirements. We also describe the effect of the size of the input datasets on the structure and execution profiles of these workflows. Finally, we describe a workflow generator that produces synthetic, parameterizable workflows that closely resemble the workflows that we characterize. We make these workflows available to the community to be used as benchmarks for evaluating various workflow systems and scheduling algorithms.Workflows in Support of Large-Scale Science, 2008. WORKS 2008. Third Workshop on; 12/2008
Chapter: Workflows for e-Science[Show abstract] [Hide abstract]
ABSTRACT: Workflows for e-Science is divided into four parts, which represent four broad but distinct areas of scientific workflows. In the first part, Background, we introduce the concept of scientific workflows and set the scene by describing how they differ from their business workflow counterpart. In Part II, Application and User Perspective, we provide a number of scientific examples that currently use workflowsfalse, 01/2007: pages 1-8; Springer London., ISBN: 978-1-84628-519-6
Fine-Grained Profiling for Data-Intensive Workflows
∗‡Department of Computer Science,†Department of Information and Communication Engineering
Graduate School of Information Science and Technology, The University of Tokyo
7-3-1 Hongo, Bunkyo-ku, Tokyo, 113-8656 Japan
Abstract—Profiling is an effective dynamic analysis approach
to investigate complex applications. ParaTrac is a user-level pro-
filer using file system and process tracing techniques for data-
intensive workflow applications. In two respects ParaTrac helps
users refine the orchestration of workflows. First, the profiles of
I/O characteristics enable users to quickly identify bottlenecks
of underlying I/O subsystems. Second, ParaTrac can exploit
fine-grained data-processes interactions in workflow execution
to help users understand, characterize, and manage realistic
data-intensive workflows. Experiments on thoroughly profiling
Montage workflow demonstrate that ParaTrac is scalable to
tracing events of thousands of processes and effective in guiding
fine-grained workflow scheduling or workflow management
Keywords-profile, workflow, file system trace
Recent advances in the cluster, grid, and cloud computing
enable users to execute various data-intensive workflows
by harnessing widely available computing resources .
However, planning and scheduling the execution of complex
workflows in distributed environments still remain chal-
lenging . One of important demands is to understand
and characterize the realistic behaviors of data-intensive
workflows to help workflow management systems refine
their orchestration for optimal workflow execution.
In response to this practical demand, research has been
conducted to elaborate the characterization of a wide variety
of scientific workflows using synthetic approaches , .
Though these methods are capable of understanding the
basic structures of representative workflows from domain
researchers, they show a lack of capturing fine-grained data-
processes activities and it will be ideal if users can acquire
the workflow essentials from their own workflows.
Profiling is an effective dynamic analysis approach to
investigate complex applications in practice. Accordingly,
we designed and implemented a user-level profiler ParaTrac
for data-intensive workflow applications using file system
and process tracing techniques. First, ParaTrac can produce
valuable file system call statistics and input/output charac-
terizations for I/O subsystem (e.g., distributed file system)
performance analysis and tuning. Second, ParaTrac is novel
because it provides realistic and comprehensive profiles of
data-intensive workflows by exploiting fine-grained data-
process interactions. These informative essentials enable
users to understand and characterize the workflows, and
therefore refine the planning, scheduling, and execution of
complex workflows in distributed environments.
II. PROFILING APPROACHES
Tracing an application and producing profiles by ParaTrac
are straightforward and effortless. First, using ParaTrac, user
creates a traced file system via which application can be
executed without modification simply by changing working
directory or using chroot utility. User-level techniques used
to trace both files and processes activities of unmodified ex-
ecutables includes FUSE (Filesystem in Userspace), /proc
file system, and kernel process accounting.After the appli-
cation finished, all system call events and process activities
are stored into raw time-series logs for further processing.
Then, the trace logs produced from individual trace instances
are collected and integrated into one global SQL-queryable
database. Finally, using the profile generation utility, the
profiles of application are created by synthetically mining
the trace data in database.
For I/O characterization profiles, statistical analysis is
applied to reveal the overall behaviors of data manipulation,
such as the variances of system calls and the I/O regularity
and sequentiality of workloads.
For workflow analysis, ParaTrac uses casual and temporal
analysis to generate process hierarchical tree (i.e. fork
graph) as well as the workflow DAG (Directed Acyclic
Graph) annotated with fine-grained runtime statistics to
illustrate actual data-processes interactions in workflow ex-
ecution. For workflows that are executed in distributed
environments, the DAGs are aggregated from trace logs
produced at each host. Then various manipulations can be
applied to DAGs for further analysis. For example, the
analysis of sub-workflow is available by partitioning DAG
into subgraphs. The scheduling analysis can be achieved by
aggregating processes-level DAG into job/task level DAG.
In addition, graph-theoretic algorithms can be applied to
node and edge attributes to explore specific tasks or data in
workflow. For example, file nodes having the most degrees
may indicate the critical node in workflow.
522.00B@69.46MB/sec522.00B@69.46MB/sec 522.00B@69.46MB/sec 522.00B@69.46MB/sec
1.91KB@78.04KB/sec 1.91KB@78.04KB/sec1.91KB@78.04KB/sec 1.91KB@78.04KB/sec
Access by Subprocess
(a) Complete Workflow DAG
(b) Sub-DAG of Workflow
Figure 1. Fine-Grained DAG of Montage Workflow
Our experiments use ParaTrac to profile the execution
of Montage scientific workflow for different data sets on
8 Linux servers in the wide-area environments.
Experimental results show that ParaTrac introduces about
16% tracing overhead, which is mainly due to the context
switch in FUSE and the tracing scalability of one trace
instance mainly depends on the capability of FUSE handling
concurrent requests. The size of tracing data primarily scales
with the number of system call events, and is secondar-
ily correlated with the number of processes born during
the execution. For montage workflow, the trace data are
approximately proportional (about 10%) to the size of the
application data processed by the whole workflow.
From system call and I/O profiles, system calls with
high latency in Montage workflow can be easily discovered.
For example, operations that manipulate remote data having
much higher latency than average ones indicate a proper data
prefetch or replication strategy should be applied.
Figure 1 illustrates the remarkable feature of ParaTrac:
the DAGs of Montage workflow with very detailed data-
processes dependencies and interactions.
• Process: Processes (name#lifetime in oval) with
their parent-child relationships (dot line with ‘o’ ar-
rowhead) during the workflow execution.
• Data: Files (filename in box) accessed during the
workflow execution are annotated with the size, portion
and transfer rate (size#rate on edge) of data accessed
by and passed among processes.
Comparing to the workflow DAGs by Pegasus Workflow-
Generator , , ParaTrac provides more informative and
fine-grained DAGs. Since more realistic runtime information
can be obtained in details by profiling than synthetic analy-
sis, we suggest that workflow management systems provide
extra parameters specifying the resource constraints to allow
users to perform the fine-grained scheduling of workflows.
For example, a workflow management system can utilize the
data pass volume and transfer rates obtained in workflow
profiles to achieve a data throttling strategy to improve the
throughput of workflows.
IV. CONCLUSION AND FUTURE WORK
We have presented the design and implementation of
ParaTrac — a profiler targeted for data-intensive workflows.
Experiments on thoroughly profiling Montage workflow by
ParaTrac demonstrate its scalability of tracing events of
thousands of processes and its effectiveness by guiding fine-
grained workflow scheduling and workflow management
In the future, we plan to expand ParaTrac to trace more
workflow information, such as CPU time of tasks, to enrich
the profiles of applications. Another direction is to reuse the
profiles as a macro benchmark for workflow management
systems by consistent replaying of profiles.
ParaTrac is an open source software and online available
 I. Taylor, E. Deelman, D. B. Gannon, and M. S. (Eds),
Workflows for e-Science. Springer-Verlag, 2006.
 S. Pandey and R. Buyaa, “Scheduling and management tech-
niques for data-intensive appliation workflows,” in Data In-
tensive Distributed Computing: Challenges and Solutions for
Large-Scale Information Management, T. Kosar, Ed.
 S. Bharathi, A. Chervenak, E. Deelman, G. Mehta, M.-H. Su,
and K. Vahi, “Characterization of scientific workflows,” in
Proc. The 3rd Workshop on Workflows in Support of Large-
Scale Science, 2008, pp. 1–10.
 L. Ramakrishnan and D. Ganno, “A survey of distributed
workflow characteristics and resource requirements,” Indiana
University, Tech. Rep. TR671.
 Workflow generator. [Online]. Available: http://vtcpc.isi.edu/