|
|
|
| e-Pub |
Previous | 
Home
Section: New Results
Experimental methodologies for the evaluation of distributed systems
This year, M. Quinson defended his Habilitation on the experimental methodologies of distributed systems [13] . This concludes 10 years of research on this topic (including the elements presented in this section), and paves the road of future research.
Simulation and dynamic verification
MPI simulation
Participants : Martin Quinson, Paul Bédaride, Marion Guthmuller.
We continued our long-term effort toward the simulation of HPC application within SimGrid. We slightly increased the API coverage of our reimplementation of MPI on top of SimGrid, and proposed a new model of the network performance for MPI applications on top of Ethernet TCP networks. This model combines the advantages of flow-based networks for large data transfers as previous SimGrid network models, but also leverage algorithmic performance models extending the classical LogP models. As shown in [16] , these models greatly improve the realism of MPI simulations, enabling the prediction of the performance of a non-trivial application in great details.
Dynamic verification and SimGrid
Participants : Marion Guthmuller, Martin Quinson, Gabriel Corona.
This year, our work toward the verification of liveness properties within SimGrid became fully functional thanks to the PhD work of M. Guthmuller. This relies on a system-level introspection mechanism allowing the model checker to finely explore the state of the verified programs. This is mandatory to detect the execution cycles that constitute the counter examples to liveness properties. This introspection mechanism is also used to implement a new reduction mechanism that can mitigate the state space explosion problem. A publication presenting these results is currently under review.
SimGrid framework improvement
Participants : Paul Bédaride, Martin Quinson, Gabriel Corona.
We rolled out a new major version of the SimGrid framework to our users. It contains both the HPC network models used to improve the prediction of MPI applications and all of our developments toward the dynamic verification of distributed applications. We also improved further the usability of our framework, that is now properly integrated within the Debian Linux distribution.
The next release is already underway, with a proper integration of the work from our partners on virtual machines and with a full reimplementation of the simulation kernel in C++ for a better modularity.
Formal Verification of Distributed Algorithms
Participants : Esteban Campostrini, Martin Quinson, Stephan Merz.
M. Quinson co-advised an internship with S. Merz (project-team Veridis) on the formal verification of distributed algorithm. The goal was to push further the PlusCal algorithmic language and its compiler to TLA on which we are working since several years within the Veridis team.
We wanted to explore some hard problem raised by the verification of distributed protocol, such as how to represent timeout errors in verification settings where the time is not present. We think that this could be modeled somehow similarly to fairness properties, but more work is needed in this topic for a definitive answer.
Experimentation on testbeds and production facilities, emulation
Distem improvements: scalability and matrix-based inter-nodes latencies
Participants : Ahmed Bessifi, Emmanuel Jeanvoine, Lucas Nussbaum.
(For context, see sections 3.3 and 5.3 .)
Following our PDP'13 publication[18] , we focused on improving Distem's scalability. First, on the Distem engine side, we parallelized the startup of physical nodes and virtual nodes, and added support for BTRFS snapshots to enable starting a very large number of virtual nodes with their own filesystems. Second, during the internship of Ahmed Bessifi we investigated several networking issues causing problems with large-scale experiments (over 4000 virtual nodes). The resulting improvements to ARP parameters tunings were integrated in Distem 0.8, and enabled network-intensive experiments with up to 8000 virtual nodes. We plan to publish those results in early 2014.
In the context of the AEN HEMERA project, we worked with Trong-Tuan Vu (EPI DOLPHIN, Inria Lille Nord Europe) to add support for specifying inter-nodes latencies using a matrix. This is especially useful for experiments on load-balancing and locality.
Evaluating load balancing HPC runtimes with Distem
Participants : Joseph Emeras, Emmanuel Jeanvoine, Lucas Nussbaum.
(For context, see sections 3.3 and 5.3 .)
We aim at demonstrating the suitability of Distem to evaluate Exascale and Cloud runtime environments providing load balancing and fault tolerance features. In that context, we reproduced some experiments published at CCGrid'2013 on Charm++ load balancers. Preliminary results are promising, and we hope that this will lead to collaborations with runtime developers.
A publication presenting how Distem to test HPC runtimes (scalability, fault tolerance and load balancing capabilities) is in the works.
Further improvements to XPFlow
Participants : Tomasz Buchert, Lucas Nussbaum, Jens Gustedt.
(For context, see sections 3.3 and 5.6 .)
We strengthened our XPFlow experiment control system using several sets of experiments, including experiments on the OpenStack IaaS Cloud stack on hundreds of Grid'5000 nodes.
A publication describing XPFlow was submitted to CCGrid'2014[21] .
Further improvements to Kadeploy
Participants : Luc Sarzyniec, Emmanuel Jeanvoine, Lucas Nussbaum.
(For context, see sections 3.3 and 5.5 .)
We continued the development of Kadeploy:
-
The communication interface between the Kadeploy server and the Kadeploy client was completely rewritten to use a REST API;
-
A test framework, integrated with Inria's Continuous Integration facility, was added.
Two new Kadeploy releases were published during 2013, including those changes.
Grid'5000
Participants : Sébastien Badia, Luc Sarzyniec, Émile Morel, Lucas Nussbaum.
(For context, see sections 3.3 and 5.7 .)
The team continued to support Grid'5000. Highlights of 2013 include:
-
Lucas Nussbaum is now a member of the Bureau and Comité d'Architectes of GIS Grid'5000. In the context of the Comité d'Architectes, he led the writing on several internal documents (on possible evolutions of the testbed).
-
An article describing Grid'5000's support for experiments on IaaS Clouds[15] was published at the Testing The Cloud workshop.
Convergence and co-design of experimental methodologies
Practical study on combining experimental methodologies
Participants : Maximiliano Geier, Lucas Nussbaum, Martin Quinson.
During an internship, we explored how simulation, emulation and experimentation on Grid'5000 could be combined in practice. Starting with a simple question on a particular system, we used a representative framework for each methodology: SimGrid for simulation, Distem for emulation and Grid’5000 for experimentation, and described our experiments using the workflow logic provided by the XPFlow tool. We identified a set of pitfalls in each paradigm that experimenters may encounter regarding models, platform descriptions and others. We proposed a set of general guidelines to avoid these pitfalls. We showed these guidelines may lead to accurate simulation results. Finally, we provided some insight to framework developers in order to improve the tools and thus facilitate this convergence.
The results of this work were published at the WATERS workshop[17] .
Organization of an event on reproducible research
Participant : Lucas Nussbaum.
We organized Realis, an event aiming at testing the experimental reproducibility of papers submitted to Compas'2013. Associated to the Compas'13 conference, this workshop aimed at providing a place to discuss the reproducibility of the experiments underlying the publications submitted to the main conference. We hope that this kind of venue will motivate the researchers to further detail their experimental methodology, ultimately allowing others to reproduce their experiments.