| EQUIPE
ASSOCIEE |
MODSIM (MODular SIMulation) |
|
sélectionnée en |
|
|
sélectionnée en |
|
|
sélectionnée en |
|
|
soumise en |
2005 |
| Projet INRIA : ALCHEMY | Organisme étranger partenaire : Liberty Group, Computer Science Department, Princeton University |
| Unité de recherche INRIA : Futurs Thème INRIA : Systèmes communicants |
Pays : Etats-Unis |
|
Coordinateur
français |
Coordinateur
étranger | |
| Nom, prénom | Temam, Olivier | August, David |
| Grade/statut | Directeur de Recherche | Assistant Professor |
| Organisme d'appartenance |
INRIA Futurs | Computer Science Department, Princeton University |
| Adresse postale | Parc Club Orsay Université ZAC des vignes 4, rue Jacques Monod - Bât G 91893 Orsay Cedex France |
Computer Science Building, Room 209 35 Olden Street Princeton University Princeton NJ, 08544 Etats-Unis |
| URL | Alchemy group | http://www.cs.princeton.edu/~august |
| Téléphone | +33 1 72 92 59 52 | +1 609 258 20 85 |
| Télécopie | +33 1 60 19 66 08 | +1 609 258 17 71 |
| Courriel | olivier.temam@inria.fr | august@cs.princeton.edu |
Proposal Summary
|
Keywords : |
| Collaboration Summary : Processor architecture
simulation is the key, if not the exclusive, method used by
micro-architecture researchers for evaluating the performance and
usefulness of new architecture ideas. Micro-Architecture simulation often
consists of writing very large software programs describing the detailed
behavior of processors, i.e., what is happening in the processor at every
clock cycle, the architecture performance being often measured in numbers
of cycles. While these simulators are not as detailed as circuit-level
models, the software complexity is very much tied to the architecture
complexity itself, and with high-performance processors quickly evolving
from a few million transistors in the 1990s to a billion transistors or
more in 2005, simulators are becoming excessively complex pieces of
software, which are very time-consuming to develop and modify.
Consequently, a single research group can no longer afford to develop a whole new processor simulator as part of its normal research activities, i.e., to evaluate a few architecture ideas. As a result, most researchers rely on available models, and especially SimpleScalar, developed at University of Wisconsin in 1997, and the most heavily used by far (in 50% to 70% of the articles in top architecture conferences). Now, SimpleScalar and most other simulators have a serious drawback: they are monolithic, i.e., they are built as a single piece of software describing a full processor architecture; but a processor architecture is typically composed of several tens of components (cache, scheduler, branch predictions...), which have evolved differently over the years, and more importantly, researchers are usually specialized in just one or a few of these components. And because simulators such as SimpleScalar do not reflect the modularity of processor architectures, it can be exceedingly difficult to update one or a few architecture components, or to extract a component proposed and implemented by a researcher in order to compare it with other similar components. This situation has severe consequences on micro-architecture research: (1) researchers develop their ideas on outdated architecture models because they lack alternatives, (2) simulator development time is quickly increasing and researchers have no simple means for reusing and exploiting the development effort of other researchers, (3) researchers have no reasonably rigorous way to compare their performance results, and consequently, these results are often unverifiable and unreliable, (4) the discrepancy between the software structure and the processor structure is a source of inaccuracy sometimes leading researchers to propose unrealistic architectures, (5) and as result, it hinders the take-up of architecture ideas from academia by industry. In the past few years, the Liberty group at Princeton and the INRIA Alchemy group have both been working on an alternative approach to processor simulation called modular simulation (respectively the Liberty project and the MicroLib project), the basic principle being to reflect the processor structure in the software structure, alleviating many of the above mentioned issues. Both groups share the goal of changing the methodology in processor architecture research by encouraging architecture researchers to use modular simulation environments, and by setting up a central repository where researchers could easily exchange, reuse and compare architecture ideas through simulator components. The rationale behind this partnership proposal is to collaborate rather than compete in this task, and to propose a common modular simulation approach, thereby increasing the rate at which the community will adopt a modular methodology. Recent successes registered by both groups suggests this goal, while ambitious, may not be unrealistic. The Liberty group is the leading group in modular simulation in the US, and the INRIA Alchemy group is similarly trying to promote this approach in Europe, especially as part of the HiPEAC Network of Excellence; thus, this partnership is a way to set up a first bridge between Europe and the US on this topic, ultimately creating a universal standard. |
Presentation of the Partnership
1. Presentation
of the foreign coordinator
Prof. David August
(PDF CV) is
an Assistant Professor in the Department of Computer Science at
Princeton University. His research interests lie in computer
architecture and back-end compilation. At Princeton, he directs the
Liberty Group. The Liberty
group is currently developing open-source tools necessary to perform
rigorous, systematic processor design-space exploration. Liberty
research has been sponsored in part by Intel, IBM, Infineon, The
Gigascale Silicon Research Center, an NSF NGS, an NSF ITR, and an NSF
CAREER Award. David's work has been recognized by an NSF CAREER
Award, an Emerson Electric Company - E. Lawrence Keyes '51 Award, and
two best paper awards. In 2001, he was invited by University of
Virginia to participate to the Top Gun Lecture Series
recognizing "faculty on a trajectory to be research leaders of the
coming decades". And in 2005, he will be the General Chair of the next
International Symposium on Code Generation and Optimization
(CGO). David received the B.S. (Summa Cum Laude) in Electrical
Engineering from Rensselaer Polytechnic Institute in 1993, and the
M.S. and Ph.D. degrees in Electrical and Computer Engineering from the
University of Illinois at Urbana-Champaign in 1995 and 2000,
respectively (University of Illinois is regularly ranked as one of the
top Electrical and Computer Engineering schools in the US). At
Illinois, as a member of the IMPACT research compiler group, he
developed a complete framework for aggressive predicate analysis and
optimization currently available as part of the OpenIMPACT
toolset.
The Liberty Research Group has been, along with the INRIA Alchemy group, a recognized developer, proponent, and leader of the modular and structural simulation movement. In 2002, the Liberty Group published a paper making the case for modular and structural simulation at the top conference MICRO. This early paper demonstrated that the current monolithic simulator practice is both costly and error prone. It also demonstrated that modular simulation could provide an attractive solution. For their work, the authors won a best student paper award at this top-tier computer architecture conference.
Since that time, the Liberty Research Group has developed prototype tools to demonstrate the reality of modular simulation. They released the Liberty Simulation Environment (LSE) version 1.0 to the community to facilitate early modular and structural simulation work (over 800 downloads). Tutorials at two top-tier computer architecture conferences (ASPLOS and MICRO) have been given to potential users. Other publications relating to the language of LSE and its optimizing simulator-building compiler have also been published in respected conferences (PLDI-2004, CODES+ISSS-2004, DAC-2003).
Using LSE, the Liberty Research group and others have demonstrated the effectiveness of structural simulation in a variety of computer architecture research work. Using structural modeling methods, one student was able to create a validated Itanium 2 model accurate to within 3% in only 11 weeks. This compares favorably to the person-years it generally takes to get to within 10-20% error. In a testament to the reuse and ease of modification of structural models, this validated model was used to create a multiprocessor model in only 16 days. This multiprocessor model incorporated a new interconnection mechanism presented by the Liberty group at PACT in Antibes Juan-les-Pins this past October. The validated Itanium 2 model has also been used to study fault tolerant design techniques and server workloads in heterogeneous multiprocessor systems. Other models developed include a PowerPC model, an extremely accurate validated multi-code network interface controller model (The SPINACH project led by Prof. Vijay Pai at Purdue University), and a grid processor model with an accurate power model (The ORION project led by Prof. Li-Shiuan Peh at Princeton University).
While the successes of LSE are many, this early work and experience raised more research questions and revealed inefficiencies that we hope to address in this joint collaboration.
The Liberty Research Group has been, along with the INRIA Alchemy group, a recognized developer, proponent, and leader of the modular and structural simulation movement. In 2002, the Liberty Group published a paper making the case for modular and structural simulation at the top conference MICRO. This early paper demonstrated that the current monolithic simulator practice is both costly and error prone. It also demonstrated that modular simulation could provide an attractive solution. For their work, the authors won a best student paper award at this top-tier computer architecture conference.
Since that time, the Liberty Research Group has developed prototype tools to demonstrate the reality of modular simulation. They released the Liberty Simulation Environment (LSE) version 1.0 to the community to facilitate early modular and structural simulation work (over 800 downloads). Tutorials at two top-tier computer architecture conferences (ASPLOS and MICRO) have been given to potential users. Other publications relating to the language of LSE and its optimizing simulator-building compiler have also been published in respected conferences (PLDI-2004, CODES+ISSS-2004, DAC-2003).
Using LSE, the Liberty Research group and others have demonstrated the effectiveness of structural simulation in a variety of computer architecture research work. Using structural modeling methods, one student was able to create a validated Itanium 2 model accurate to within 3% in only 11 weeks. This compares favorably to the person-years it generally takes to get to within 10-20% error. In a testament to the reuse and ease of modification of structural models, this validated model was used to create a multiprocessor model in only 16 days. This multiprocessor model incorporated a new interconnection mechanism presented by the Liberty group at PACT in Antibes Juan-les-Pins this past October. The validated Itanium 2 model has also been used to study fault tolerant design techniques and server workloads in heterogeneous multiprocessor systems. Other models developed include a PowerPC model, an extremely accurate validated multi-code network interface controller model (The SPINACH project led by Prof. Vijay Pai at Purdue University), and a grid processor model with an accurate power model (The ORION project led by Prof. Li-Shiuan Peh at Princeton University).
While the successes of LSE are many, this early work and experience raised more research questions and revealed inefficiencies that we hope to address in this joint collaboration.
2. History of the collaboration
-
2.1. Between the research groups :
Scientific collaboration. The scientific collaboration includes several aspects: currently, defining a common modular simulation approach, and later, make the appropriate software developments and further demonstrate the benefits of modular simulation through publications and other activities.
- Communication protocol. Currently, the collaboration is
focused on defining a common modular simulation approach,
especially a common communication protocol between simulation
modules (a module is a simulator of a processor architecture
component). The key point is to avoid multiple standards and to
ascertain that future modules, whether developed at Princeton,
INRIA, or anywhere else will be interoperable so the community
will not have to choose between different standards. While the
main principles of both approaches are similar, Liberty and ModLib
modules cannot yet communicate because of protocol and software
reasons. In the past two months, we have been working on a common
communication protocol that would realize a good tradeoff between
ease of programming (what the user is exposed to and has to do),
and speed (how fast will the simulation execute). Speed is a key
issue because simulators typically execute 10000 to 100000 times
slower than an actual architecture, meaning one second of
execution on a 2GHz complex processor requires between 3 and 30
hours of simulation on a fast workstation. At the moment, each
environment behaves best in certain simulation cases and we feel
it is possible to design a protocol that retains both assets. At
the same time, it is just as important that the user who develops
a simulation module feels it is easy to convert any architectural
idea into a model and grasps/takes advantage of the benefits of
modular simulation (more rigorous architecture modeling, easy
reuse of modules,...). For that part, while the Liberty approach
may be more adequate for synchronous processors (the vast majority
of current processors), we are investigating what would be an
appropriate approach for asynchronous processors (a possibly
emerging class of low-power architectures, as recently proposed by
ARM).
- Capabilities. The common communication
protocol is a first step toward an integrated modular simulation
approach. Once interoperability is achieved, the second step is to
understand what features users want to find in a simulator, how to
integrate them in a simulator, and how these features may evolve;
these features are called capabilities. Examples of such
features are power & temperature modeling, layout, graphical
visualization of simulation, etc. For the approach and the
corresponding developments to be long-lasting, we need to find
both a generic approach to integrating such capabilities, as well
as immediately integrate a number of key capabilities in the first
simulators.
- Central repository. Comparison and
reuse of ideas/efforts are key benefits of modular simulation. We
believe they will become a reality only if we can provide a
central repository where researchers will get and put simulator
components. The INRIA group has focused on showing the benefits of
such a central repository through the MicroLib project, and we
need to unify the method for sharing/disseminating modular
simulator components. Ultimately, the repository will enable
researchers to publish their modules and characterize their
simulators as a set of repository modules in their articles. With
this, other researchers will be able to easily reproduce the
experiments, something which is impossible today even though it is
commonplace in more mature scientific domains, such as biology and
physics. This repository raises new issues of architecture
interoperability (in addition to protocol & software
interoperability mentioned above). It also brings new benefits for
architecture research, as explained in the
workplan.
As we now reach a phase of the collaboration where more technical discussions and joint developments are necessary, we need funding to ensure regular meetings and stays of PhDs and faculties at each site for a few days to several weeks.
As of now, there are no joint publications because the collaboration is recent, but the collaboration has been increasingly active, and we intend to submit a joint article making a case for modular simulation in early 2005, and we do intend to have additional publications in the future. Moreover, the nature of the project is such that it will most likely breed joint development first, and later on joint publications. Joint development started in June 2004 when Daniel Gracia-Perez was at Princeton. During the course of this development the scope of the collaboration has increased, requiring even more scientific interactions.
Recent actions and achievements. As mentioned in the abstract, both groups have registered several scientific and adoption successes. The Liberty group has first published its modular simulation approach in the top-level conference MICRO in 2002, and received the best student paper award; it has also published in 2004 a presentation of its modular modeling language in another top-level conference (PLDI). The group has also been very active in promoting the modular simulation approach through multiple tutorials and seminars, and several research groups are already using their environment; Intel, which has been exploring a modular approach for its own simulators in the past few years, has been evaluating the Liberty environment and is contemplating adopting this approach.
The INRIA Alchemy group has set up its MicroLib web site in 2002, and has focused on populating its Library with full processor models and modular simulation components; one of the first models disseminated on the web site has already been downloaded 2500+ times. The models have already been used in several publications, as well as for system test at EADS and CEA. Recently, the MicroLib project has illustrated the benefits of modular simulation for comparing research ideas through an article to appear at MICRO 2004. This work has attracted the attention of Intel, and it has asked the MicroLib group to organize the second international challenge on micro-architecture in 2005.
Finally, in September 2004, the HiPEAC network has expressed interest in the approach of the Liberty and Alchemy groups, and will be contemplating a joint proposal as a candidate for its common simulation framework, confirming that it viewed a joint Europe-US collaboration as the best way to maximize acceptance chances. However, again, the HiPEAC NoE is not funding this partnership and does not plan to in the future as it is beyond its role. - Communication protocol. Currently, the collaboration is
focused on defining a common modular simulation approach,
especially a common communication protocol between simulation
modules (a module is a simulator of a processor architecture
component). The key point is to avoid multiple standards and to
ascertain that future modules, whether developed at Princeton,
INRIA, or anywhere else will be interoperable so the community
will not have to choose between different standards. While the
main principles of both approaches are similar, Liberty and ModLib
modules cannot yet communicate because of protocol and software
reasons. In the past two months, we have been working on a common
communication protocol that would realize a good tradeoff between
ease of programming (what the user is exposed to and has to do),
and speed (how fast will the simulation execute). Speed is a key
issue because simulators typically execute 10000 to 100000 times
slower than an actual architecture, meaning one second of
execution on a 2GHz complex processor requires between 3 and 30
hours of simulation on a fast workstation. At the moment, each
environment behaves best in certain simulation cases and we feel
it is possible to design a protocol that retains both assets. At
the same time, it is just as important that the user who develops
a simulation module feels it is easy to convert any architectural
idea into a model and grasps/takes advantage of the benefits of
modular simulation (more rigorous architecture modeling, easy
reuse of modules,...). For that part, while the Liberty approach
may be more adequate for synchronous processors (the vast majority
of current processors), we are investigating what would be an
appropriate approach for asynchronous processors (a possibly
emerging class of low-power architectures, as recently proposed by
ARM).
-
2.2. between INRIA and the foreign institution :
We are not aware of other collaborations between INRIA and Princeton University.
- Liberty group at Princeton University
- Computer Science Department at Princeton University
- MicroLib project at INRIA
- INRIA DREI site for United States
- CNRS DRI site for United States
- French embassy in the United States
3. Impact :
First, it must be noted that a characteristic of this
collaboration is that its scientific goal is fairly clear, and that it may have
a significant scientific impact on our research domain as explained above, as well as on the take-up of ideas
from academia by industry; both
the timing and the synergies are right. So this proposal should not only be
evaluated based on the usual criteria of the benefits for a collaboration among
two institutions (INRIA & Princeton), but also based on its scientific
impact and timeliness, as explained above.
- 3.1. (on the existing collaboration with your
partner)
As explained above, the collaboration has been increasingly active and is currently limited solely by the lack of meeting opportunities. We are not requesting funding to start a collaboration, we are requesting funding to make our collaboration successful, to achieve the goals we have set by stepping up the pace of interaction. Joint developments cannot be resolved through phone conversations, people have to meet in person for extended periods of time and regularly. Moreover, we are trying to merge our two approaches, so we need to expose people in each group to the approach of the other group, and that requires stays in each group.
- 3.2. (on other INRIA research groups)
The INRIA CAPS group, headed by Andre Seznec, is very active in micro-architecture research and has already expressed interest in using our simulation methodology. So, while they will not take part to the research activities described in this proposal, they will directly benefit from a standard approach to modular simulation, and the corresponding tools.
- 3.3. (on other research groups at the foreign
institution)
Several research groups at Princeton University have expressed an interest in using the joint system. Professors Kai Li and Li-Shuian Peh lead research groups that have benefited from the Liberty Simulation Environment. This collaboration would strengthen the value of modular simulation for them and others at Princeton as it would make more tools available for them and speed their research. This is the same impact it could have on the entire international computer architecture research community.
4. Miscellaneous :
II. BILAN 2005
|
Eventuelles remarques et/ou changements survenus (indiquez ici, le cas échéant, les éléments des années antérieures qui vous semblent importants ):
|
Bilan synthétique des 3
dernières années
|
Rapport scientifique pour l'année 2005
Description de l'activité scientifique de l'équipe associée et des résultats obtenus : publications, communications, organisation de colloques, formation, soutenance de thèse, valorisation économique, sociale, industrielle, dépôt de brevets ... (1 à 2 pages)