OpenIB Developers Workshop, February 6-9, 2005, Sonoma, CA

Presentations

Full Sessions

Opening pdf
Speaker: Bill Boas - LLNL
Introduction: Welcome, Workshop Team, Workshop Overview, Goals, Desired Results in OpenIB Alliance, Q&A.

HPC Keynote pdf
Speaker: Mark Seager - LLNL
Abstract:
Requirements for interconnect in today's 100 Tflop machines and transition to PetaFlop architectures.

OpenIB Directions pdf
Speaker: Roland Dreier - Topspin
Abstract:
What it took to get kernel acceptance and where can we go from here.

Enterprise Keynote pdf
Speakers: Ty Panagoplos and Peter Krey - JP Morgan

Quality Assurance C and I, Testing, and Release
Moderator: Jim Ryan - Intel
Speaker: Tuan Phamdo - Intel pdf
Abstract:
This presentation will explain C&I as it exists as a major focus area for the IBTA, and describe the current organization. It will then propose CIWG and OpenIB can have a complementary relationship by focusing on interoperability for OpenIB, but drawing from a foundation of expertise and best practices from the CIWG.

Speaker: Richard Hedges - LLNL Scalable I/O pdf
Abstract:
This presentation will focus on the process applied at LLNL to introduce a new technology and roll it out to our customers in a process that allows them to be productive. Aspects of the process include the organization of the team supporting the new technology, extensive in house testing focused upon user requirements, controlled internal release processes, and collaboration from end users in debugging. We draw from our most recent experiences with the Lustre file system and project our strategy for InfiniBand integration into new clusters.

Quality Assurace Panel
Moderator: Jim Ryan
Panel:Matt Leininger - Sandia; Richard Frank - Oracle; Richard Hedges - LLNL, Tuan Phamdo - Intel
Abstract:
The QA panel will discuss both what is needed and what is available to take the OpenIB SW stack to the highest quality levels possible. The National Labs speaker will describe testing facilities available there and how to take advantage of them. The Oracle speaker will describe SW testing suites, how to get them and the sequence of steps to go through to use them to get certification. The speakers from JP Morgan Chase (JPMC) will describe their quality requirements and testing for their own cluster development and high-availability application support.

OpenIB Plans for 2005 Moderators: Jim Ryan - Intel, Matt Leininger - Sandia Panel: Chris Eddington - Mellanox, Mike Boorman - LANL, Asaf Samekh - Voltaire, Phil Murphy - InfiniCon, Richard Frank - Oracle, Boris Bialek - IBM
Abstract:
During the Workshop the panel members will be capturing the suggestions/agreements contemplated by the participants. The panel will discuss these and engage the attendees in reprising the coding intentions, quality assurance, release, distribution, industry development activities and collaborations the the Alliance will focus on in 2005.

Developers Track

D1 - Core Software 1 pdf pdf
Speakers: Hal Rosenstock - Voltaire, Roland Dreier - Topspin, Sean Hefty - Intel
Abstract:
This session describes the architecture and implementation of the existing core software (access layer) that was submitted to kernel.org. This includes the mthca driver, the core access layer, mad services, CM and the SA query module.

D2 - Core Software 2 pdf
Abstract:
This session is a continuation D1 and discusses the user-mode access infrastructure and additional capabilities that were not implemented in the initial patch that was submitted to LKML. It will also outline the Diagnostic tools and utilities that are under development, such as tvflash, and the diagnostic tools planned for implementation, including ibstatus, ibping, ibroute, ibtracert, smaquery, perfquery.

D3 - OpenSM pdf
Speaker: Eitan Zahavi - Mellanox
Abstract:
This session describes the OpenSM subnet manager and its capabilities.

D4 - uDAPL/kDAPL pdf pdf
Speaker: Arkady Kanevsky and James Lentini- Network Appliance
Abstract:
This session describes the RDMA Direct Access Provider Library (DAPL) being developed by the DAT collaborative.

D5,6,and 7 - MPI Sessions
Moderator: Curt Janssen - Sandia
Speakers:
Dr. Panda (OSU) - OSU MPI MVAPICH pdf
Mitch Sukalski (Sandia) - LA-MPI pdf and OpenMPI pdf
Bob Woodruff (Intel) - Intel MPI pdf
Greg Lindahl (Pathscale) - MPI ABI pdf
Abstract:
These three session describes the versions of MPI that are be developed to utilize the new OpenIB Stack.

D8 - IPoIB and SDP pdf
Speaker: Libor Michalek - Topspin
Abstract:
This session describes the existing IPoIB driver and the architecture
of the sockets direct protocol driver that is currently under development.

D9 - iSER (iSCSI Extension for RDMA) pdf
Speaker: Yaron Haviv, Voltaire
Abstract:
iSER is a new IETF extension to iSCSI allowing it to run over native RDMA transport replacing the slow TCP transport. iSER utilizes the same control and management interfaces that iSCSI does. This way it combines the reach, functionality and scalability of iSCSI with the performance of InfiniBand. The session will describe the new iSCSI Datamover layer, the iSER transport module and its relations with iSCSI and InfiniBand, the standard mapping from iSCSI/iSER to FC and to iSCSI/TCP. The session will also review the progress in incorporating iSER into the iSCSI-Linux project, the latest iSER RFC work in the IETF and iSER integration with OpenIB.

D10 - SRP
Speaker: Tony Vaidya - Topspin
Abstract:
This session will present the SCSI RDMA protocol (SRP) and plans for porting it to the new openib stack.

D11 - Panel: Scalability in Hardware and Software for HPC
Moderator: Matt Leininger - Sandia
Speakers:
Ron Minnich (LANL) - LANL Clusters pdf
Steve Poole (LANL) - HPC Requirements for IB Hardware
Michael Kagan (Mellanox) - Mellanox Chip Directions pdf

D12 - IBM Linux Technology Directions
Speaker: Venkata Jagana
Abstract:
The subject topic provides an overview of IBM's Linux Technology Center, its role in Open Source projects as well as past contributions to in the Linux IB effort. It also details IBM's Linux IB needs, current priorities and future areas of contribution to OpenIB.

D13 - InfiniBand Fabric, Beyond Microseconds and Megabytes pdf
Speaker: Yaron Haviv - Voltaire
Abstract:
InfiniBand Architecture was designed to address the new age of commercial clustering and resource consolidation, as such it spans to additional essential capabilities such as Security, Partitioning, QoS, centralized addressing and routing, performance monitoring, high-availability, and congestion management. The session will cover how the above capabilities are implemented in InfiniBand, and how they can be exploited to optimize cluster operation, improve scalability, and/or to save costs. In addition the session will discuss possible ways for OpenIB software to make optimal use of these capabilities.

D14 - IBTA verbs 1.2 and other verbs extensions pdf
Speaker: Dror Goldenberg - Mellanox
Abstract:
IB spec 1.2 introduces new verbs extensions that optimize performance, scalability, and improve application security. The session goes through the verbs extensions and covers the functionality, application benefits and driver interfaces.

D15 - OpenIB Verification Requirements pdf
Speaker: Hal Rosenstock - Voltaire
Abstract:
This is a discussion of the verification and test needs for OpenIB from a release and production standpoint from the developers perspective. Verification needs to start at each layer/component and build from there. The talk is is intended to lay the ground work for Wednesday's sessions by proposing a shared set of goals for a verification architecture and an open set of test tools, test procedures and problem reporting centered around the OpenIB repository and web site.

Application and Industry Development Track

A1 - InfiniBand Connected Mass Storage pdf
Speaker: Dave Fellinger - DataDirect Networks
Abstract:
Cluster architectures have evolved with various network layers and protocols. Low latency high cost switches have been used to tie the compute cluster together while block level storage interfaces included SCSI, Fibre Channel, and Ethernet through servers or gateway nodes. Infiniband allows all of the elements of the cluster to be tied together by a common interface and common network. Data transfer protocols taking full advantage of the RDMA capabilities will allow individual compute nodes to more efficiently transfer data to and from storage devices. Storage architectures will be explored that allow data transfers directly to a file system memory space in any node without data copies. Potential features of a system with a shared memory map will also be discussed in light of file system optimization.

A2 - DB2 Database Clustering on InfiniBand Fabrics pdf
Speaker: Boris Bailek - IBM
Abstract:
This session describes how the DB2 architectures exploits InfiniBand RDMA fabrics to optimize performance and scalability. Successful deployments will be used as examples to illustrate the requirements for OpenIB in the enterprise markets. The session will close with a outlook from the DB2 team how we see the InfiniBand technology evolve for the business.

A3 - Native InfiniBand Storage Systems using SRP and iSER protocols pdf
Speaker: Bret Weber - Engenio Information Technologies, Inc.
Abstract:
This session focuses on how users can leverage InfiniBand as a storage connection to provide direct access to storage subsystems utilizing SCSI derivative protocols. It will examine some high level differences between SRP and iSER, and how storage management utilities and failover make the transition to the InfiniBand environment.

A4 - Infiniband in the Embedded Market pdf
Speaker: Todd Bowman, SBS Technologies
Abstract:
This session presents an overview of InfiniBand applications in the embedded market and the Open IB requirements specific to this market in the areas of QoS, failover, reliability and management.

A5 - kDAPL and NFS over RDMA Protocol Applications pdf
Speaker: Charles Antonelli, Univ of Michigan
Abstract:
This session will present an overview of the NFS/RDMA Linux project at the Center for Information Technology Integration. We will examine the relevant aspects of the NFSv4.1 specification, the overall architecture of the Linux server and client, our use of RDMA in the NFS RPC layer, and requirements on the kDAPL interface. Current project status and future plans will be included.

A6 - Oracle's Planned Usage and Issues pdf
Speaker: Richard Frank - Oracle
Abstract:
This session will present an overview of Oracle's use of InfiniBand in the database, and Oracle's current efforts regarding SDP, IPoIB, uDAPL, IT-API, and SRP. This session will also discuss the issues we see in supporting these ULPs in OpenIB.

A7 - Interconnect Software Consortium Update pdf
Speaker: Carl Hensler - Sun Microsystems
Abstract:
This session will provide an update on the work being done by the Interconnect Software Consortium (ICSC) including their work on the RDMA transport (IT-API) and the RDMA socket extensions for Unix.

A8 - Visualization Experience and Requirements pdf
Speaker: Mike Houston - Stanford University Graphics Lab
Abstract:
Visualization clusters have unique network requirements. This session discuss the design of current graphics cluster, the clusters recently built at several DOE facilities, as well as new prototype clusters in development. How parallel rendering is distributed across a graphics cluster using packages like Chromium will be described, including several rendering configurations and their network usage characteristics. Chromium has a NAL which is currently mapped to an existing IB stack with which we have had issues and there are features that need to added. The talk will conclude with new research into computation on clusters of GPUs, and their networking requirements.

A9 - IBTA and OpenIB Developing Relationship pdf
Speaker: Allyson Klein - Intel and Chris Eddington - Mellanox
Abstract:
IBTA and OpenIB are complimentary organizations that have a common interest in InfiniBand architecture's commercial adoption. This session discusses the technical and promotional activities where the two organizations can potentially collaborate.

A10 - MemFree HCA pdf
Speaker: Dror Goldenberg - Mellanox
Abstract:
InfiniHost III-Ex introduces the Mem-Free Technology. The HCA no longer requires attached memory, instead the HCA uses host memory to store its context tables. The session describes the HW/SW modifications, initial performance experience and optimizations.

A11 - IBM IB Trends and Directions pdf
Speaker: Chet Mehta
Abstract:
The presentation provides a high level overview of IBM's InfiniBand plans along with the targeted segments. In addition it outlines IBM's view of key requirements that would accelerate IB adoption and position it for continued growth.

A12 - Server Baseboard Management over InfiniBand
Speakers:
Ron Minnich (LANL)
Adrian Wu (Appro) and Chris Eddington (Mellanox) pdf
Abstract:
There are significant advantages and interest in using InfiniBand as the only connection on the server node, however this requires that server baseboard and remote booting capability be done over the InfiniBand network. This session will discuss the requirements for remotely managing server over the InfiniBand fabric and discuss approaches for practical implementation.

A13 - Parallel Programming and Infiniband - When is one extra layer, one too many? pdf
Speaker: Grant Grundler - HP
Abstract:
HP's MPI parallel programming support library and CFS's Lustre File System are apllications that layer on IB. The problems they have, and are working to overcome, are also inhibiting other applications that could substantially expand the market for IB and the use of parallel programming paradigms in general. This talk explores these problems and solution approaches in proprietary interconnects for guidance in future IB development.

A14 - NASA Columbia System pdf
Speaker: Jeff Becker - NASA
Abstract:
TBD