International Research
Workshop on
Advanced High Performance Computing Systems
Cetraro (
Workshop Agenda
Monday, June 27th
|
Session |
Time |
Speaker/Activity |
|
|
|
Welcome Address |
|
|
Grid and Cloud Computing 1 |
|
|
|
|
V.
GETOV Smart
Cloud Computing: Autonomy, Intelligence and Adaptation |
|
|
|
P.
MARTIN |
|
|
|
J.L.
LUCAS A Multi-cloud Management Architecture and Early
Experiences |
|
|
|
COFFEE BREAK |
|
|
|
D.
PETCU How
is built a mosaic of Clouds |
|
|
|
M.
KUNZE Towards
High Performance Cloud Computing (HPCC) |
|
|
|
R.
MAINIERI The
future of cloud computing and its impact on transforming industries |
|
|
|
CONCLUDING REMARKS |
|
|
Grid and Cloud Computing 2 |
|
|
|
|
B.
SOTOMAYOR Reliable
File Transfers with Globus Online |
|
|
|
K.
MIURA RENKEI:
A Light-weight Grid Middleware for e-Science Community |
|
|
|
COFFEE BREAK |
|
|
|
P.
KACSUK Supporting
Scientific and Web-2 Communities by Desktop Grids |
|
|
|
L.
LEFEVRE Energy efficiency from networks to large scale
distributed systems |
|
|
|
M.
STILLWELL Dynamic
Fractional Resource Scheduling |
|
|
|
CONCLUDING REMARKS |
Tuesday, June 28th
Wednesday, June 29th
|
Session |
Time |
Speaker/Activity |
|
|
Advanced software issues for top scale HPC |
|
|
|
|
E.
LAURE CRESTA
- Collaborative Research into Exascale Systemware, Tools and Applications |
|
|
|
T.
LIPPERT Amdahl hits the Exascale |
|
|
|
S.
DOSANJH On
the Path to Exascale |
|
|
|
C.
SIMMENDINGER Petascale in CFD |
|
|
|
COFFEE BREAK |
|
|
Advanced Infrastructures, Projects and
Applications |
|
|
|
|
PANEL DISCUSSION Exascale
Computing: from utopia to reality |
|
|
|
CONCLUDING REMARKS |
INVITED SPEAKERS
|
R. Ansaloni |
Cray |
|
|
A. Benoit |
|
|
|
G. Bosilca |
|
|
|
E. D’Hollander |
|
|
|
S. Dosanjh |
SANDIA
National Laboratories |
|
|
V. Getov |
|
|
|
P. Kacsuk |
MTA
SZTAKI |
|
|
H. Kaiser |
|
|
|
M. Kunze |
Karlsruhe
Institute of Technology |
|
|
E. Laure |
Royal |
|
|
L. Lefevre |
INRIA RESO – LIP |
France |
|
T. Lippert |
Juelich
Supercomputing Centre |
|
|
J.L.
Lucas |
|
|
|
R. Mainieri |
IBM Italy |
|
|
P. Martin |
Queen’s
University, |
|
|
L. Mirtaheri |
|
|
|
K. Miura |
Center
for Grid Research and Development National Institute of Informatics |
Japan |
|
C. Perez |
INRIA –
LIP |
France |
|
D. Petcu |
Research
Institute e-Austria |
|
|
F. Pinel |
|
|
|
A. Shafarenko |
|
|
|
M. Shekhalishahi |
|
|
|
C. Simmendinger |
T-Systems
Solutions for Research GmbH |
|
|
B. Sotomayor |
Computation
Institute, |
|
|
L. Sousa |
INESC and TU Lisbon |
|
|
M.
Stillwell |
INRIA, |
|
ABSTRACTS
R. Ansaloni
Cray’s Approach to Heterogeneous Computing
There seems to be a general consensus among the
HPC community, about the impossibility to reach hexascale
performance with systems based only on multi-core chips. Heterogeneous nodes
where the traditional CPU is combined with many-core accelerators have the
potential to provide a much more energy efficient solution capable to overcome
the power consumption challenge.
However this emerging hybrid node architecture
is expected to pose significant challenges for applications developers, in
order to efficiently program these systems and achieve a significant fraction
of the available peak performance.
This is certainly the case for today’s
GPU-based accelerators with separate memory space, but it also holds true for
future unified nodes with CPU and many-core accelerator on chip sharing common
memory.
In this talk I’ll describe Cray’s approach to
heterogeneous computing and the first Cray hybrid supercomputing system with
its unified programming environment.
I’ll also describe Cray’s proposal to extend
the OpenMP standard to support a wide range of
accelerators.
Energy-aware mappings
of series-parallel workflows onto chip multiprocessors
In this talk, we will study the problem of mapping
streaming applications that can be odelled by a
series-parallel graph, onto a 2-dimensional tiled CMP architecture. The
objective of the mapping is to minimize the energy consumption, using dynamic
voltage scaling techniques, while maintaining a given level of performance,
reflected by the rate of processing the data streams. This mapping problem
turns out to be NP-hard, but we identify simpler instances, whose optimal
solution can be computed by a dynamic programming algorithm in polynomial time.
Several heuristics are proposed to tackle the general problem, building upon
the theoretical results. Finally, we assess the performance of the heuristics
through a set of comprehensive simulations.
G. Bosilca
Flexible Development
of Dense Linear Algebra Algorithms on Heterogeneous Parallel Architectures with
DAGuE
In the context of dense linear algebra
developing algorithms that seamlessly scales to thousands of cores can be
achieved using DPLASMA (Distributed PLASMA). DPLASMA take advantage of a novel
generic distributed Direct Acyclic Graph Engine (DAGuE).
The engine has been designed for fine granularity tasks and thus it enables
scaling of tile algorithms, originating in PLASMA, on large distributed memory
systems. The underlying DAGuE framework has many
appealing features when considering distributed-memory platforms with
heterogeneous multicore nodes: DAG representation
that is independent of the problem-size, automatic extraction of the
communication from the dependencies, overlapping of communication and
computation, task prioritization, and architecture-aware scheduling and
management of tasks.
E. D’Hollander
High-performance
computing for low-power systems
Intelligent low-power devices such as portable
phones, tablet computers, embedded systems and sensor networks require
low-power solutions for high-performance applications. GPUs
have a highly parallel multithreaded architecture and an efficient programming
model, but are power-hungry. On the other hand field programmable gate arrays
have a highly configurable parallel architecture and a substantially better
energy efficiency, but are difficult to program. An approach is presented which
maps the GPU architecture and programming model onto the configuration
synthesis and the programming of FPGAs.
Implementation details, benefits and trade-offs are discussed. In particular
the architecture, memory and communication issues are addressed when porting a
biomedical image application with a 20-fold GPU speedup onto an FPGA
accelerator.
S. Dosanjh
On the path to Exascale
This presentation will describe technical and
programmatic progress in Exascale computing. The Exascale Initiative was included in the U.S. Department of
Energy’s budget starting in the U.S. Government’s 2012 fiscal year. Several
partnerships are forming, a number of projects have already been funded and
several co-design centers are being planned. These
co-design centers will develop applications for Exascale systems and will provide feedback to computer
companies on the impact of computer architecture changes on application
performance. An enabling technology for these efforts is the Structural
Simulation Toolkit (SST), which allows hardware/software co-simulation. Another
key aspect of this work is the development of mini-applications. One difficulty
of co-design in high performance computing (HPC) is the complexity of HPC
applications, many of which have millions of lines of code. Mini-applications,
which are typically one-thousand lines of code, have the potential to reduce
the complexity of co-design by a factor of one-thousand. Mini-applications
representative of finite-elements, molecular dynamics, contact algorithms, and
shock physics are described. The performance of these mini-applications on
different computer systems is compared to the performance of the full
application.
V. Getov
Smart Cloud Computing:
Autonomy, Intelligence and Adaptation
In recent years, cloud computing has rapidly
emerged as a widely accepted computing paradigm. The cloud computing paradigm emerged
shortly after the introduction of the “invisible” grid concepts. The research
and development community has quickly reached consensus on the core cloud
properties such as on-demand computing resources, elastic scaling, elimination
of up-front capital and operational expenses, and establishing a pay-as-you-go
business model for computing and information technology services. With the
widespread adoption of virtualization, service-oriented architectures, and
utility computing, there is also consensus on the enabling technologies
necessary to support this new consumption and delivery model for information
technology services. Additionally, the need to meet quality-of-service
requirements and service-level agreements, including security, is well understood.
Important limitations of the current cloud computing systems include lack of
sufficient autonomy and intelligence based on the existence of dynamic
non-functional properties. Such properties together with support for adaptation
can change completely the quality of computerised services provided by the
future cloud systems. In this presentation, we plan to address these issues and
demonstrate the significant advantages provided to the users by the smart cloud
computing platforms. Some of the available directions for future work are also
discussed.
P. Kacsuk
Supporting scientific and Web-2 communities by desktop grids
Although the nature of scientific and Web-2
communities are different they both require more and more processing power to
run compute-intensive applications for the sake of community members.
Scientific communities typically require run large parameter sweep based
simulations that are ideal for both volunteer and institutional desktop grids.
Web-2 communities use community portals like facebook
through which they organize their social relationship and activities. Such
activities also could include time-consuming processing, like water marking the
photos of community members.
Both communities prefer to use affordable
distributed infrastructures in order to minimize the processing cost. Such a
low-cost infrastructure could be a volunteer or institutional desktop grid. The
EU EDGI project developed technology and infrastructure to support scientific
communities by desktop grids, while the Web2Grid Hungarian national project
provides desktop grid technology and the corresponding business model for Web-2
communities.
The talk will discuss the main characteristics
of such desktop grid support and also shows the major architectural components
of the supporting architecture. The application areas and the possible business
models of using volunteer desktops will also be addressed in the talk.
H. Kaiser
ParalleX – A Cure for Scaling-Impaired
Parallel Applications
High Performance Computing is experiencing a
phase change with the challenges of programming and management of heterogeneous
multicore systems architectures and large scale
systems configurations. It is estimated that by the end of this decade Exaflops computing systems requiring hundreds of millions
of cores demanding multi-billion-way parallelism with a power budgets of
50Gflops/watt may emerge. At the same time, there are many scaling-challenged
applications that although taking many weeks to complete, cannot scale even to
a thousand cores using conventional distributed programming models. This talk
describes an experimental methodology, ParalleX, that
addresses these challenges through a change in the fundamental model of
parallel computation from that of the communicating sequential processes (e.g.
MPI) to an innovative synthesis of concepts involving message-driven work-queue
computation in the context of a global address space. We will present early but
promising results of tests using a proof-of-concept runtime system
implementation guiding future work towards full scale parallel programming.
M. Kunze
Towards High
Performance Cloud Computing (HPCC)
Today’s HPC clusters are typically operated and
administrated by a single organization. Demand is fluctuating, however,
resulting in periods where dedicated resources are either underutilized or
overloaded. A cloud-based Infrastructure-as-a-Service (IaaS)
model for HPC promises cost savings and more flexibility, as it allows to move away from physically owned and potentially
underutilized HPC clusters to virtualized and elastic HPC resources available
on-demand from consolidated large cloud computing providers.
The talk discusses specific issues of the
introduction of a resource virtualization layer in HPC environments such as latency,
jitter and performance.
E. Laure
CRESTA - Collaborative
Research into Exascale Systemware,
Tools and Applications
For the past thirty years, the need for ever
greater supercomputer performance has driven the development of many computing
technologies which have subsequently been exploited in the mass market.
Delivering an exaflop (or 10^18 calculations per
second) by the end of this decade is the challenge that the supercomputing
community worldwide has set itself. The Collaborative Research into Exascale Systemware, Tools and
Applications project (CRESTA) brings together four of Europe’s leading
supercomputing centres, with one of the world’s major equipment vendors, two of
Europe’s leading programming tools providers and six application and problem
owners to explore how the exaflop challenge can be
met.
CRESTA focuses on the use of six applications
with exascale potential and uses them as co-design
vehicles to develop: the development environment, algorithms and libraries,
user tools, and the underpinning and cross-cutting technologies required to
support the execution of applications at the exascale.
The applications represented in CRESTA have been chosen as a representative
sample from across the supercomputing domain including: biomolecular
systems, fusion energy, the virtual physiological human, numerical weather
prediction and engineering.
No one organisation, be they a hardware or
software vendor or service provider can deliver the necessary range of
technological innovations required to enable computing at the exascale. This is recognised through the on-going work of
the International Exascale Software Project and, in
In this talk we will give an overview of
CRESTA, outline the challenges we face in reaching exascale
performance and how CRESTA intends to respond to them.
L. Lefevre
Energy efficiency from
networks to large scale distributed systems
Energy efficiency begins to be largely adressed for distributed systems like Grids, Clouds or
networks. These large-scale distributed systems need an ever-increasing amount
of energy and urgently
require effective and scalable solutions to manage and limit their electrical
consumption.
The challenge is to coordinate all low-level
improvements at the middleware level to improve the energy efficiency of the
overall systems. Resource-management solutions can indeed benefit from a
broader view to pool the resources and to share them according to the needs of
each user. During, this talk I will describe some solutions adopted for large
scale monitoring of distributed infrastructures. This talk will present our
work on energy efficient approaches for reservation based large scale
distributed systems. I will present the ERIDIS model, an Energy-efficient
Reservation Infrastructure for large-scale DIstributed
Systems which provides a unified and generic framework to manage resources from
Grids, Clouds and dedicated networks in an energy-efficient way.
T. Lippert
Amdahl hits the Exascale
With the advent of Petascale
supercomputers the scalability of scientific application codes on such systems
becomes a most pressing issue. The current world record holder as far as the
number of concurrent cores is concerned, the IBM Blue Gene /P system
"JUGENE" at the Jülich Supercomputing
Centre with 294 912 cores will soon be displaced by systems comprising millions
of cores. In this talk I am going to review the constraints put on scalability
by Amdahl’s and Gustafson’s Laws. I am proposing architectural concepts that
are optimized for the concurrency hierarchies of application codes and I will
give a glimpse on the DEEP Exascale supercomputer
project, to be funded by the European Community, that explicitly addresses
concurrency hierarchies on the hardware, system software and application
software level.
J.L. Lucas
A Multi-cloud
Management Architecture and Early Experiences
In this talk we present a
cloud broker architecture for deploying virtual infrastructures across
multiple IaaS clouds. We analyse the main challenges
of the brokering problem in multi-cloud environments, and we propose different
scheduling policies, based on several criteria, that can guide the brokering
decisions. Moreover, we present some preliminary results to prove the benefits
of this broker architecture in multi-cloud environments in the execution of
virtualized computing clusters.
R. Mainieri
The future of cloud
computing and its impact on transforming industries
In its centennial IBM demonstrated that
long-term success requires vision, strategy and managing for the long term.
Deciding how and where investing and allocating resources, shaping talent
development and taking decisive action. Three years ago IBM started talking
about smarter planet and how it was driving innovation across industries. On a
Smarter Planet, successful companies think differently about computing and
realize IT infrastructure that is designed for data, tuned to the task, and
managed in the cloud.
The talk will illustrate IBM cloud computing
vision, strategy and management plan for the long term: the smarter computing
for a smarter planet. It will discuss about resources invested in research and
development, present the main important global projects and how some specific
actions such as laboratories around the world, new cloud data center, software companies acquisitions, fostering the
adoption of open standards is going to lead a sustainable industry
transformation in specific industries.
P. Martin
Provisioning Data-Intensive Workloads in the Cloud
Data-intensive workloads involve significant amounts of data access. The
individual requests composing these workloads can vary from complex processing
of large data sets, such as in business analytics and OLAP workloads, to small
transactions randomly accessing individual records within the large data sets,
such as in OLTP workloads. In the cloud, applications generating these workloads
may be built on different frameworks from shared-nothing database management
systems to MapReduce or even some mix of the two. We
believe that effective provisioning methods for data-intensive workloads in the
cloud must consider where to place the data in the cloud when they are
allocating resources to the workloads.
In the talk, I will provide an overview of an approach that provisions a
workload in a public cloud while simultaneously placing the data in an optimal
configuration for the workload. We solve this data placement problem by solving
two subproblems, namely how to first partition the
data to suit the workload and then how to allocate data partitions to virtual
machines in the cloud.
L. Mirtaheri
An Algebraic Model for
a Runtime High Performance Computing Systems Reconfiguration
Tailored High Performance Computing Systems
(HPCS) represent the best performance because their configuration is customized
regarding the features of the problem to be solved. 21th
century processes are dynamic in nature. However, this dynamicity in nature is
caused either because of the dimensions of today’s problems being undetermined
or the dynamicity of the underlying platform. A drawback of this dynamicity is
for the systems customized at design phase facing challenges at runtime and
consequently showing worse performance. The reason for these challenges might
be for the processes with dynamic nature being in the opposite direction as
that of the system configuration. Many approaches like dynamic reconfiguration
with dynamic load balancing are introduced to solve the challenges. In this
talk, I will present a mathematical model based on vector algebra for system
reconfiguration. This model determines the element (process) causing the
opposition and discovers the reason of that regarding both software and
hardware at runtime. Results of the presented model show that by defining a
general status vector whose direction is towards reaching high performance and
size is based on the initial features and explicit requirements of the problem
and also by defining a vector for each process in the problem at runtime, we
can trace changes in the directions and find out the reason, as well.
K. Miura
RENKEI: A Light-weight
Grid Middleware for e-Science Community
The “RENKEI (Resources Linkages for e-Science)
Project” started in September 2008 under the auspices of the Ministry of
Education, Culture, Sports, Science and Technology (MEXT). In this project, a
new light-weight grid middleware and software tools are developed, in order to
provide the user-friendly connection between the major grid environment and
users’ local computing environment. In particular, technology for enabling the
flexible and seamless accesses to the national computing center
level and the departmental/laboratory level resources, such as computers,
storage and databases, is one of the key objectives. Another key ingredient of
this project is “interoperability” with the major international grids along the
line of OGF standardization activities, such as GIN, PGI, SAGA and RNS.
With the RENKEI workflow tool users can submit
jobs from the local environment or even from a cloud to the “TSUBAME2
supercomputer system at the Tokyo Institute of Technology, via the networking
infrastructure called “SINET4”, for example..
http://www.naregi.org/index_e.html
C.
Perez
Resource management system
for complex and non-predictably evolving applications
High-performance scientific applications are
becoming increasingly complex, in particular because of the coupling of
parallel codes. This results in applications having a complex structure, characterized
by multiple deploy-time parameters, such as the number of processes of each
code. In order to optimize the performance of these applications, the
parameters have to be carefully chosen, a process which is highly resource
dependent. Moreover, some applications are (non-predictably) changing their
resource requirements during their execution.
Abstractions provided by current
Resource Management Systems (RMS) appears insufficient to efficiently select resources
for such applications. This talks will discuss CooRM, an RMS architecture to support such applications. It
will also show how applications can benefit from it to achieve a more efficient
resource usage.
D. Petcu
How is built a mosaic
of Clouds
The developers of Cloud compliant application
are facing the dilemma of which Cloud provider API to select knowing that later
on this decision will be lead to a provider
dependence. mOSAIC (www.mosaic-cloud.eu) is addressing this issue by proposing
a vendor and language-independent API for developing Cloud compliant
applications. Moreover it has promise to build a Platform as a Service solution
that will allow the selection at run-time of the Cloud services from multiple
offers based on semantic processing and agent technologies.
The presentation will focus on the problems
raised by implementing the Sky computing concept (cluster of Clouds), the
issues of Virtual Cluster deployment on top of multiple Clouds, and the
technical solutions that were adopted by mOSAIC.
F. Pinel
Utilizing GPUs to Solve Large Instances of the Tasks Mapping Problem
In this work, we present and analyze a local
search algorithm designed to solve large instances of the independent tasks
mapping problem. The genesis of the algorithm is the sensitivity analysis of a
cellular genetic algorithm, which illustrates the benefits of such an analysis
for algorithmic design activities.
Moreover, to solve instances of up to 65,536
tasks over 2,048 machines and to achieve scalability, the local search is
accelerated by utilizing a GPU. The proposed local search algorithm improves
the results of other well-known algorithms in the modern literature.
A. Shafarenko
New-Age Component
Linking: Compilers Must Speak Constraints
This presentation will focus on the agenda of
the FP7 project ADVANCE. The project is seeking to redefine the concept of
component technology by investigating the possibility of exporting out of a
component not only interfaces, but functional and extrafunctional
constraints as well. The new, rich component interface requires a hardware
model for the aggregation and resolution of constraints, but if that is
available, then a much more targeted approach can be defined for compiling
distributed applications down to heterogeneous architectures.
Constraint aggregation can deliver the missing
global (program-wide) intelligence to a component compiler and enable it to
tune up the code for alternative hardware, communication harness or memory
model.
The talk will discuss these ideas in some
detail and provide a sketch of a Constraint Aggregation Language, developed in
the project.
M. Sheikhalishahi
Resource Management
and Green Computing
In this talk, we review green and performance
aspects of resource management. Components of resource management system are
explored in detail to seek new developments by exploiting contemporary emerging
technologies, computing paradigms, energy efficient operations, etc. to define,
design and develop new metrics, techniques, mechanisms, models, policies, and
algorithms. In addition, modeling relationships
within and between various layers are considered to present some novel
approaches. In particular, as a case study we define and model resource
contention metric and consequently we develop two energy aware consolidation
policies.
C. Simmendinger
Petascale in CFD
In this talk we outline a highly scalable and
also highly efficient PGAS implementation for the CFD solver TAU.
TAU is an unstructured RANS CFD Solver and one of
the key applications in the European aerospace Eco-System. We show that our
implementation is able to scale to petascale systems
within the constraints of a single regular production run.
To reach this goal, we have implemented a novel
approach for shared memory parallelization, which is based on an asynchronous
thread pool model. Due to the asynchronous operation the model is implicitly
load balanced, free of global barriers and also allows for a near-optimal
overlap of communication and computation. We have complemented this model with
an asynchronous global communication strategy, in which we made use of the PGAS
API of GPI.
We briefly outline this strategy and show first
results.
B. Sotomayor
Reliable File Transfers with Globus Online
File transfer is both a critical and
frustrating aspect of high-performance computing. For a relatively mundane
task, moving terabytes of data reliably and efficiently can be surprisingly
complicated. One must discover endpoints, determine available protocols, negotiate
firewalls, configure software, manage space, negotiate authentication,
configure protocols, detect and respond to failures, determine expected and
actual performance, identify, diagnose and correct network misconfigurations,
integrate with file systems, and a host of other things. Automating these makes users’ lives much, much easier.
In this presentation I will provide a technical
overview of Globus Online: a fast, reliable file
transfer service that simplifies large-scale, secure data movement without
requiring construction of custom end-to-end systems. The presentation will
include a demonstration as well as highlights from several user case studies.
L. Sousa
Distributed computing
on highly heterogeneous systems
The approaches used in traditional
heterogeneous distributed computing to achieve efficient execution across a set
of architecturally similar compute nodes (such as CPU-only distributed systems)
are only partially applicable to the systems with a high degree of
architectural heterogeneity. For example, when considering clusters of
multi-core CPUs equipped with specialized accelerators/co-processors, such as GPUs. This is mainly due to the fact that efficient load
balancing decisions must also be made at the level of each compute node, in addition
to the decisions made at the overall system level.
In this work, we propose a method for dynamic
load balancing and performance modeling for
heterogeneous distributed systems, when all available compute nodes and all
devices in compute nodes are employed for collaborative execution. Contrary to
the common practice in task scheduling, we do not make any pre-execution
assumptions to ease the modeling of either the
application or the system. The heterogeneous system is modeled
as it is, by monitoring and recording the behavior of
the essential parts affecting the performance.
The parallel execution requires explicit data
transfers to be performed prior to and after any actual computation. In order
to exploit the concurrency between data transfers and computation, we
investigate herein the processing in an iterative multi-installment
divisible-load space at both, overall system and compute node levels. Namely,
the proposed approach dispatches the load using many sub-loads, whose size is
carefully determined to allow the best overlap between communication and
computation. The load division is performed according to several factors: i) current performance models (per-device and per-node),
ii) modeled bidirectional interconnection bandwidths
(between compute nodes and between devices in each compute node), and iii) the
amount of supported concurrency by the node/device hardware.
The problem that we tackle herein is how to
find task distribution, such that the overall application make-span is the
shortest possible according to the current performance models of devices,
interconnections and compute nodes. Performance models are application-centric
piece-wise linear approximations constructed during the application runtime to
direct further load-balancing decisions according to the exact task
requirements.
The proposed approach is evaluated in a real
distributed environment consisting of quad-core CPU+GPU nodes, for iterative
scientific applications, such as matrix multiplication (DGEMM), and Fast
Fourier 2D batch Transform (FFT). Due to ability to overlap execution of
several sub-loads, our approach results in more accurate performance models
comparing to the current state-of-the-art approaches.
M. Stillwell
Dynamic Fractional
Resource Scheduling
Dynamic Fractional Resource Scheduling is a
novel approach for scheduling jobs on cluster computing platforms. Its key
feature is the use of virtual machine technology to share \emph{fractional}
node resources in a precise and controlled manner. Our previous work focused on
the development of task placement and resource allocation heuristics to
maximize an objective metric correlated with job performance, and our results
were based on simulation experiments run against real traces and established
models. We are currently performing a new round of experiments using synthetic
workloads that launch parallel benchmark applications in multiple virtual
machine instances on a real cluster. Our goals are to see how well our ideas
work in practice and determine how they can be improved, and to develop
empirically validated models of the interaction between resource allocation
decisions and application performance.