Performance characterization of computational resources for time-constrained job execution in P2P environments

Research output not available from this repository.

Request-a-Copy directly from author or use local Library Get it For Me service.

Request Changes to record.

Abstract

Peer-to-peer (P2P) computing, involving the participation of thousands of general
purpose, public computers, has established itself as a viable paradigm for executing looselycoupled,
complex scientific applications requiring significant computational resources. The
paradigm provides cheap, general-purpose computing resources with comparable computing
power (FLOP/s) to an otherwise expensive supercomputer. The main characteristic of the
paradigm is the volunteer participation of the general public, without any legal obligation,
who dedicate their heterogeneous computational resources for advancing scientific research.
The development of several middleware solutions have also furthered the application of P2P
computing for solving complex scientific problems. The Berkeley Open Infrastructure for
Network Computing (BOINC) is one of the most widely deployed middleware platforms in
P2P systems, and has been deployed in more than 7.5 million general purpose computers for
scientific computations, achieving an overall performance of 16,632.605 TeraFLOPS.
ClimatePrediction.net, a large P2P project based on the BOINC middleware, involves more
than 429,000 machines representing 200 different microprocessor architectures and running
21 distinct operating systems. The availability of such a large and diverse set of
computational resources requires an in-depth investigation into the performance aspects of
available computational resources in this dynamic P2P environment.
This thesis analyses the performance data of ClimatePrediction.net primarily collected
using two benchmarks, Dhrystone and Whetstone, which form part of the BOINC
middleware. The results reveal a significant variation in integer and floating-point
operational performance characterized by Dhrystone and Whetstone respectively for similar
microprocessors, operating systems and hardware configurations. Under the BOINC
environment, these performance results could be useful for: i) the selection of a suitable
computing platform for executing time-constrained jobs; ii) calculating an incentive unit for
rewarding project participants for their volunteer participation in large P2P projects to
advance scientific research; and iii) efficient and effective utilization of available
computational resources. However, the inconsistency in performance results of Dhrystone and Whetstone significantly affect their usefulness for the afore-mentioned three important
applications areas, and highlight the need for reliability and consistency of performance
results for obtaining maximum benefit in an uncontrolled and dynamic P2P environment.
This thesis, based on the analysis of performance data of ClimatePrediction.net, identifies
the key challenges associated with benchmarking in P2P environments. The thesis further
suggests the design of a new light-weight P2P representative benchmark, by considering the
source code of large P2P projects. The design outline of a new light-weight P2P
representative benchmark – MalikStone – has been presented, whilst the results of
MalikStone are compared with Dhrystone, Whetstone and CPU SPEC2006 and show its
superiority in terms of consistency over both Dhrystone and Whetstone. For floating-point
performance, MalikStone gave more representative results than Whetstone for Intel Corei5-
2400, Q9400, Q6600 and Pentium D processors with the standard deviation of repeated runs
remaining less than 1 for each of the platforms. Similarly for integer operations, MalikStone
also performed more consistently than Dhrystone with the standard deviation of repeated
runs remaining less than 1 and gave more representative results for Corei5-2400, Q9400,
Q6600 and Pentium D processors. In addition to the consistency in performance results,
MalikStone captures broader performance characteristics by measuring floating-point,
integer, bitwise-logic, string manipulation and programming construct operations.
The performance results of MalikStone are further used for designing a new incentive
unit – MalikCredit – for ensuring fairness in rewarding the project participants for their
volunteer participation in large P2P projects to advance scientific research. MalikCredit is
compared with BOINC’s existing incentive unit – Cobblestone, at three levels: 1) hourly
level; 2) work-unit level; and 3) team-level; with the results showing fairness in rewards
awarded using MalikCredit. This in turn is useful for retaining the existing project
participants and attracting new volunteers for participating in large P2P projects, thereby,
enhancing the application of P2P computing for solving scientific problems. A comparison
of the credit values for the considered microprocessor architectures reveals that MalikCredit
values are at least 2X more than Cobblestone values before normalization while the
difference increases up to 3.3X for the fastest microprocessor, once normalization is applied
to the claimed Cobblestones.
The application of performance characterization done by MalikStone is further extended
for scheduling computational resources by dynamically slicing the work-units keeping in
view the available computational time of the resources and estimated execution time of the
work-unit. The results of this new scheduling policy highlight their usefulness in maximizing the utilization of available computational resources when compared to
BOINC’s traditional scheduling policies. The results have revealed that the policy improved
the utilization of available computational resources by approximately 10% for the
considered set of computational resources under the experimental setup considered in the
case study (see Chapter 5).
The findings of this thesis are envisaged to be primarily of significance to three main
stakeholders: i) application developers; ii) project participants; and iii) project
administrators. For application developers, the performance characterization done by
MalikStone will be useful in exploiting the characteristics of underlying platforms for
efficient execution, while at the same time supporting the improvement efforts for future
versions of the software. The results will support project participants by informing them as
to the amount of RAM, swap memory and main memory consumed during execution. The
fairness in received rewards will encourage the existing project participants to continue
participating in the lengthy execution of large P2P projects and will motivate the new
volunteers to dedicate their computational resources to join large P2P projects. For the
project administrators, the findings of this thesis will be useful in identifying suitable
processor, operating system and hardware component configuration for best-case execution.
In such a case the middleware might be instructed to postpone the allocation of work until a
more effective architecture became available. Further, the newly proposed scheduling policy
involving dynamic slicing of work-units based on the performance characterization of
MalikStone could be deployed for improving the utilization of available computational
resources.
Finally, a few avenues of future research have been identified, which if explored could
further enhance the appeal of this dynamic and uncontrolled P2P computing paradigm for
cheaply solving complex and lengthy scientific problems that otherwise require enormous
amount of financial cost as well as computational resources even exceeding that of
traditional supercomputers.

Request changes or add full text files to a record

Repository staff actions (login required)

View Item