Open Fabrics BOF Supercomputing 2008 Tziporet Koren, Gilad Shainer, Yiftah Shahar, Bob Woodruff, Betsy Zeller Rev. 0.9
2 Agenda Introduction – 5 minutes OFA objectives and goals Open Fabrics Linux Update (15 – minutes) OFED 1.3, 1.3.1, 1.4 releases OFED 1.5 plans and roadmap Open Fabrics Windows Update (15 – minutes) WinOF 2.0 release WinOF 2.1 plans and roadmap Open Discussion – 25 minutes
3 Open Fabrics BOF - Introduction Open Fabrics “The mission of the OpenFabrics Alliance (OFA) is to develop, distribute and promote a unified, transport-independent, open-source software stack for RDMA-capable fabrics and networks, including InfiniBand and Ethernet “ Support for both Microsoft Windows and Linux Development happens using a typical open source development model Developer’s Working Group code developed and reviewed using lists Validation and release by the Enterprise Working Group and the Windows Working Group OFED – for Linux, WinOF – for Windows,
4 Open Fabrics BOF - Introduction Open Fabrics – working groups (cont.) XWG – Executive Working group overall steering committee, IWG – Interoperability Working Group Open Fabrics logo program, LWG – Legal Working Group handles code licensing and other legal issues, MWG – Marketing Working Group Promotes the work of the alliance and recruits new members UWG – Users Working Group define end user requirements
5 Open Fabrics BOF - Introduction Open Fabrics – End User Lists (general mail list for all end users)
6 Agenda Introduction – 5 minutes OFA objectives and goals Open Fabrics Linux Update (15 – minutes) OFED 1.3, 1.3.1, 1.4 releases OFED 1.5 plans and roadmap Open Fabrics Windows Update (15 – minutes) WinOF 2.0 release WinOF 2.1 plans and roadmap Open Discussion – 25 minutes
7 Linux: OFED Components HCA/NIC Drivers IB: IBM, Mellanox, QLogic iWARP: Chelsio, NetEffect (now Intel) Core: Verbs, mad, SMA, CMA, SA cache IPoIB SDP SRP and SRP Target iSER RDS Qlogic_VNIC uDAPL OSM Diagnostic tools OFED 1.4 additions: iSER Target NFS-RDMA Bonding module Open iSCSI MPI Components MVAPICH Open MPI MVAPICH2 Benchmark tests Proprietary MPIs: Intel, HP, Platform mpi Proprietary SMs: Cisco, Voltaire, Qlogic OFA DevelopmentAdd on Tested with
8 Update from Sonoma ’08 Session Progress: Developers try to get SW into kernel first. If not, submit to next available kernel, and pull separately into OFE. Warnings greatly reduced, also resulting in potential bugs found Interoperability “pre-test” event was run on OFED 1.4 RC2, allowing vendors to make bug fixes before GA Identified and published list of libraries which might “change without notice” was successful, and reasonably low impact on community. Met goal of not changing kernel, and accepting only bug fixes. Evaluating and tracking buglist carefully and in detail. Community members are improving at tracking their bugs. Acquired and published list of what each vendor plans to test Still improving: Enabling distros Evaluating impact of changes to existing protocols Get community “on-board” with feature freeze date. Large changes still coming in late for some features. However, on the plus side, process enabled changing on new feature to “optional” because it wasn’t quite ready. Reviewing copyrights and licenses As we provide OFED diagnostics on Windows, we need to re-think having libraries change without notice.
9 OFED 1.3 and OFED 1.3 released on Feb Improved performance and scalability Quality of Service (QOS) support New OpenSM routing algorithms IPoIB connect mode – improved performance MPI enhancements to improve scalability and reduce memory footprint Enhanced diagnostic tools to improve network management Support for additional hardware, added additional iWarp NIC drivers Expanded interoperability with Storage and Networking hardware uDAPL 2.0 OFED release on June Added support for RedHat EL 5.2 and SLES 10 SP2 fixed several critical bugs Distro integration: Red Hat AS 4.7 and RHEL 5.2, SLES10 SP2 Used in Intel ® Cluster Ready Solutions
10 OFED 1.4 General Info Released in Nov 2008 Used in the Interoperability event in Nov 2008 Features: Kernel base VPI support: Eth and IB for ConnectX NFS-RDMA – as technology preview ISer Target New BMME verbs (fast memory thru send queue (FRWR); Local invalidate send work requests; Read with invalidate) IPoIB – LRO RDS – GA with RDMA API SDP – GA uDAPL (socket CM for scalability; UD extensions)
11 OFED 1.4 (cont.) Features (cont) OpenSM: Cached routing; APM - disjoint paths; Path balancing for LMC + console diagnostics; OpenSM configuration unification; MGID to MLID mapping for IPv6 SNM; Chained routing engines; IBA additions; Failover/Handover improvements Congestion Control in ibutils MPI: MVAPICH 1.1; MVAPICH2 1.2; OpenMPI Likely Distro Integration Red Hat EL 5.4, SLES 11, SLERT11 Used in Intel ® Cluster Ready Solutions
12 OFED 1.4 OS Matrix kernel.org: kernel and Novell SLES 10 SLES 10 SP1 (up1) SLES 10 SP2 Redhat RHEL 4 (up4, up5, up6, up7) RHEL 5 (up1, up2) OEL OEL 5 Free distros (with limited QA): Open SuSE 10.3 Fedora Core 9 Ubuntu 6.06 (with RPM package installed) CentOS 5.2 * new for OFED 1.4 in bold
13 OFED 1.5 Plans Preliminary Schedule Feature Freeze: 3/20/09 Alpha Release: 3/20/09 Beta Release: 4/20/09 RC1: 5/5/09 RC2-RCx: About every 2 weeks as needed Release: June 2009 Features: Kernel.org: and Multiple Event Queues to support Multi-core CPUs NFS/RDMA – GA RDS support for iWarp OpenMPI 1.3 Add support/backports for RedHat EL 5.3 and EL 4.8 Support for Mellanox vNIC (EoIB) and FCoIB with BridgeX device more TBD…
14 OFED 1.5 OS Matrix kernel.org: kernel and Novell SLES 10 SLES 10 SP1 (up1) SLES 10 SP2 SLES 11 Redhat RHEL 4 ( up4, up5, up6, up7, u8) RHEL5 no updates, RHEL 5 (,up1, up2, u3), OEL OEL 5 Free distros (with limited QA): Open SuSE 10.3 Fedora Core 9 Ubuntu 6.06 (with RPM package installed) CentOS 5.2 new for OFED 1.5 in bold drop support for items in blue
15 Agenda Introduction – 5 minutes OFA objectives and goals Open Fabrics Linux Update (15 – minutes) OFED 1.3, 1.3.1, 1.4 releases OFED 1.5 plans and roadmap Open Fabrics Windows Update (15 – minutes) WinOF 2.0 release WinOF 2.1 plans and roadmap Open Discussion – 25 minutes
16 Windows OpenFabrics (WinOF) Collaborative effort to develop, test and release OFA software for Windows Components – Kernel and User Space Broader test participation Add-on components for vendors to differentiate above WinOF
17 Supported Platforms Architectures x86, x86_64, IA64 Operating systems Windows XP 32&64 Windows Server 2003 Windows Cluster Compute Server 2003 Windows Vista Windows Server 2008 Windows HPC Server 2008 WHQL’ed Windows Server 2003 Windows Cluster Compute Server 2003
18 WinOF Software Stack WinOF
19 WinOF Maintainers Release coordinators - Stan Smith (Intel), Ishai Rabinovitz (Mellanox) IBAL, HCA drivers - Leonid Keller (Mellanox) NetworkDirect – Leonid Keller (Mellanox) SRP – Leonid Keller (Mellanox) & Eleanor Witiak (QLogic) IPoIB – Tzachi Dar (Mellanox) & Anh Duong (QLogic) WSD – Tzachi Dar (Mellanox) OpenSM – Yevgeny Kliteynik (Mellanox) DAPL/DAT, Installer – Stan Smith (Intel) VNIC – Alex Estrin (QLogic) Winverbs, WinMad, OFED compat libraries – Sean Hefty (Intel) Qualification sites Intel, Mellanox, Microsoft, QLogic, Voltaire and others
20 WinOF 1.1 Release HCA Drivers Access Layer (IBAL) NDIS 5.1 (IPoIB) WSD provider * SRP initiator ** uDAPL OpenSM VNIC MSI installer Tools Verbs perf Benchmarks MPI Components Microsoft Intel HP SDP Cluster management ComponentsADD-ONs Components supported by individual vendors may vary * Not available on Windows XP ** Not available on Windows XP 32 bit
21 WinOF 2.0 Release Add-on Features Server 2008 & Vista support WinVerbs framework in place Will enable RDMA over InfiniBand and Ethernet Supports OFED libibverbs APIs. Network Direct Interface support Improving performance for Microsoft MPI based applications (compared to Windows CCS 2003) Performance tuning IPoIB partitioning support Checked/debug drivers available
22 Futures IPoIB Connected mode Higher performance capabilities for TCP/UDP over InfiniBand IPoIB enhanced error logging WinVerbs fully supported Support OFED librdmacm InfiniBand diagnostic tools InfiniBand boot device support All InfiniBand-based HCAs supported
23 How to Participate? Developing code Sending patches and comments to the mailing list Doing QA Opening bugs in Bugzilla ( When opening a new bug you can choose OpenFabrics WindowsOpenFabrics Windows
24 Agenda Introduction – 5 minutes OFA objectives and goals Open Fabrics Linux Update (15 – minutes) OFED 1.3, 1.3.1, 1.4 releases OFED 1.5 plans and roadmap Open Fabrics Windows Update (15 – minutes) WinOF 2.0 release WinOF 2.1 plans and roadmap Open Discussion – 25 minutes