Section: New Results

Scalable I/O and Virtualization for Exascale Systems

Damaris/Viz

Participants : Matthieu Dorier, Gabriel Antoniu, Lokman Rahmani.

In the context of the Joint Inria/UIUC/ANL Laboratory for Petascale computing (JLCP), we are developing Damaris, which enables efficient I/O, data analysis and visualization at very large scale from SMP machines. The I/O bottlenecks already present on current petascale systems as well as the amount of data written by HPC applications force to consider new approaches to get insights from running simulations. Trying to bypass the need for storage or drastically reducing the amount of data generated will be of outmost importance for exascale. In-situ visualization has therefore been proposed to run analysis and visualization tasks closer to the simulation, as it runs.

We investigated the limitations of existing in-situ visualization software and proposed Damaris/Viz, a new version of Damaris that fills the gaps of these software by providing in-situ visualization support to Damaris. The use of Damaris/Viz on top of existing visualization packages allows us to:

• Reduce code instrumentation to a minimum in existing simulations,

• Gather the capabilities of several visualization tools to offer adaptability under a unified data management interface,

• Use dedicated cores to hide the run time impact of in-situ visualization and

• Efficiently use memory through a shared-memory-based communication model.

Experiments were conducted on Blue Waters (Cray XK6 at NCSA), Intrepid (BlueGene/P at ANL) and Grid'5000 with representative visualization scenarios for the CM1  [33] atmospheric simulation and the Nek5000  [35] CFD solver. Part of these experiments were carried by NCSA researcher Roberto Sisneros, who gave us important (and very positive) feedbacks on the usability of Damaris at scale (up to 6400 cores on Blue Waters) with real applications. The results of this work were presented as a poster in the PhD forum of IEEE IPDPS 2013 [22] , published in a research report [29] and at the IEEE LDAV 2013 conference [23] , and a demo of Damaris/Viz was presented at Inria's exhibition booth at the Supercomputing (SC 2013) conference.

This work enlightened the fact that interactive in-situ visualization, although greatly improved by Damaris/Viz, still lakes interactivity. Several meetings were organized with Tom Peterka (ANL) and Roberto Sisneros (NCSA) during the SC conference and during the 10th workshop of the JLPC. We started working on an approach that leverages information theory metrics to automatically find important features of the simulations' data and to reduce the visualization load accordingly.

CALCioM

Participants : Matthieu Dorier, Gabriel Antoniu.

Unmatched computation and storage performance in new HPC systems have led to a plethora of I/O optimizations ranging from application-side collective I/O to network and disk-level request scheduling on the file system side. As we deal with ever larger machines, the interference produced by multiple applications accessing a shared parallel file system in a concurrent manner becomes a major problem. Interference often breaks single-application I/O optimizations, dramatically degrading application I/O performance and, as a result, lowering machine wide efficiency.

Following discussions initiated in 2012 with ANL's Rob Ross and Dries Kimpe, a three month internship of Matthieu Dorier at Argonne National Lab during the summer 2013 led to the design and evaluation of CALCioM (Cross-Application Layer for Coordinated I/O Management), a framework that aims to mitigate I/O interference through the dynamic selection of appropriate scheduling policies. CALCioM allows several applications running on a supercomputer to communicate and coordinate their I/O strategy in order to avoid interfering with one another. Several I/O strategies were evaluated using this framework. Experiments on Argonne's BG/P Surveyor machine and on several clusters of Grid'5000 showed how CALCioM can be used to efficiently and transparently improve the scheduling strategy between several otherwise interfering applications, given specified metrics of machine wide efficiency.

Future work will explore approaches to automatically detect the temporal I/O patterns of simulations in order to further improve the scheduling decisions made by CALCioM.

Scalable metadata management for WAN

Participants : Rohit Saxena, Alexandru Costan, Gabriel Antoniu.

BlobSeer-WAN is a data management service specifically optimized for geographically distributed environments. It is an extension of BlobSeer, a large scale data management service. The metadata is replicated asynchronously for low latency. There is a version manager on each site and vector clocks are used to enable collision detection and resolution under highly concurrent access. It was developed within the framework of Viet-Trung Tran' s PhD thesis, in relation to the FP3C project.

BlobSeer-WAN is used as a storage backend for HGMDS, a multi master metadata server designed for a global distributed file system, developed at University of Tsukuba. Several experiments have been conducted with this setup on the Grid'5000 testbed which have shown scalable metadata performance under geographically distributed environments.