用户手册目录InfiniPath User Guide1Table of Contents5Section 1 Introduction111.1 Who Should Read this Guide111.2 How this Guide is Organized111.3 Overview121.4 Switches121.5 Interoperability121.6 What’s New in this Release131.7 Supported Distributions and Kernels141.8 Software Components151.9 Conventions Used in this Document161.10 Documentation and Technical Support16Section 2 InfiniPath Cluster Administration192.1 Introduction192.2 Installed Layout192.3 Memory Footprint202.4 Configuration and Startup222.4.1 BIOS Settings222.4.2 InfiniPath Driver Startup232.4.3 InfiniPath Driver Software Configuration232.4.4 InfiniPath Driver Filesystem232.4.5 Subnet Management Agent242.4.6 Layered Ethernet Driver242.4.6.1 ipath_ether Configuration on Fedora and RHEL4252.4.6.2 ipath_ether Configuration on SUSE 9.3, SLES 9, and SLES 10262.4.7 OpenFabrics Configuration and Startup292.4.7.1 Configuring the IPoIB Network Interface292.4.8 OpenSM302.5 SRP312.6 Further Information on Configuring and Loading Drivers312.7 Starting and Stopping the InfiniPath Software312.8 Software Status332.9 Configuring ssh and sshd Using shosts.equiv332.9.1 Process Limitation with ssh352.10 Performance and Management Tips352.10.1 Remove Unneeded Services352.10.2 Disable Powersaving Features362.10.3 Balanced Processor Power372.10.4 SDP Module Parameters for Best Performance372.10.5 CPU Affinity372.10.6 Hyper-Threading382.10.7 Homogeneous Nodes382.11 Customer Acceptance Utility40Section 3 Using InfiniPath MPI433.1 InfiniPath MPI433.2 Other MPI Implementations433.3 Getting Started with MPI433.3.1 An Example C Program443.3.2 Examples Using Other Languages453.4 Configuring MPI Programs for InfiniPath MPI463.5 InfiniPath MPI Details473.5.1 Configuring for ssh Using ssh-agent473.5.2 Compiling and Linking493.5.3 To Use Another Compiler503.5.3.1 Compiler and Linker Variables513.5.4 Cross-compilation Issues513.5.5 Running MPI Programs523.5.6 The mpihosts File533.5.7 Console I/O in MPI Programs543.5.8 Environment for Node Programs543.5.8.1 Environment for Multiple Versions of InfiniPath or MPI553.5.9 Multiprocessor Nodes563.5.10 mpirun Options563.6 Using Other MPI Implementations593.7 MPI Over uDAPL593.8 MPD593.8.1 MPD Description603.8.2 Using MPD603.9 File I/O in MPI603.9.1 Linux File I/O in MPI Programs603.9.2 MPI-IO with ROMIO613.10 InfiniPath MPI and Hybrid MPI/OpenMP Applications613.11 Debugging MPI Programs623.11.1 MPI Errors623.11.2 Using Debuggers623.12 InfiniPath MPI Limitations63Appendix A Benchmark Programs65A.1 Benchmark 1: Measuring MPI Latency Between Two Nodes65A.2 Benchmark 2: Measuring MPI Bandwidth Between Two Nodes66A.3 Benchmark 3: Messaging Rate Microbenchmarks67A.4 Benchmark 4: Measuring MPI Latency in Host Rings69Appendix B Integration with a Batch Queuing System71B.1 A Batch Queuing Script71B.1.1 Allocating Resources71B.1.2 Generating the mpihosts File72B.1.3 Simple Process Management73B.1.4 Clean Termination of MPI Processes73B.2 Lock Enough Memory on Nodes When Using SLURM74Appendix C Troubleshooting75C.1 Troubleshooting InfiniPath Adapter Installation75C.1.1 Mechanical and Electrical Considerations75C.1.2 Some HTX Motherboards May Need 2 or More CPUs in Use76C.2 BIOS Settings76C.2.1 MTRR Mapping and Write Combining77C.2.2 Incorrect MTRR Mapping77C.2.3 Incorrect MTRR Mapping Causes Unexpected Low Bandwidth78C.2.4 Change Setting for Mapping Memory78C.2.5 Issue with SuperMicro H8DCE-HTe and QHT704078C.3 Software Installation Issues79C.3.1 OpenFabrics Dependencies79C.3.2 Install Warning with RHEL4U279C.3.3 mpirun Installation Requires 32-bit Support79C.3.4 Installing Newer Drivers from Other Distributions80C.3.5 Installing for Your Distribution81C.4 Kernel and Initialization Issues81C.4.1 Kernel Needs CONFIG_PCI_MSI=y82C.4.2 pci_msi_quirk82C.4.3 Driver Load Fails Due to Unsupported Kernel83C.4.4 InfiniPath Interrupts Not Working83C.4.5 OpenFabrics Load Errors If ib_ipath Driver Load Fails84C.4.6 InfiniPath ib_ipath Initialization Failure85C.4.7 MPI Job Failures Due to Initialization Problems85C.5 OpenFabrics Issues86C.5.1 Stop OpenSM Before Stopping/Restarting InfiniPath86C.5.2 Load and Configure IPoIB Before Loading SDP86C.5.3 Set $IBPATH for OpenFabrics Scripts86C.6 System Administration Troubleshooting86C.6.1 Broken Intermediate Link87C.7 Performance Issues87C.7.1 MVAPICH Performance Issues87C.8 InfiniPath MPI Troubleshooting87C.8.1 Mixed Releases of MPI RPMs87C.8.2 Cross-compilation Issues88C.8.3 Compiler/Linker Mismatch89C.8.4 Compiler Can’t Find Include, Module or Library Files89C.8.5 Compiling on Development Nodes90C.8.6 Specifying the Run-time Library Path90C.8.7 Run Time Errors With Different MPI Implementations91C.8.8 Process Limitation with ssh93C.8.9 Using MPI.mod Files93C.8.10 Extending MPI Modules94C.8.11 Lock Enough Memory on Nodes When Using a Batch Queuing System95C.8.12 Error Messages Generated by mpirun96C.8.12.1 Messages from the InfiniPath Library96C.8.12.2 MPI Messages98C.8.12.3 Driver and Link Error Messages Reported by MPI Programs101C.8.13 MPI Stats102C.9 Useful Programs and Files for Debugging103C.9.1 Check Cluster Homogeneity with ipath_checkout103C.9.2 Restarting InfiniPath103C.9.3 Summary of Useful Programs and Files104C.9.4 boardversion105C.9.5 ibstatus106C.9.6 ibv_devinfo106C.9.7 ident106C.9.8 ipath_checkout107C.9.9 ipath_control108C.9.10 ipathbug-helper109C.9.11 ipath_pkt_test109C.9.12 ipathstats109C.9.13 lsmod110C.9.14 mpirun110C.9.15 rpm110C.9.16 status_str110C.9.17 strings112C.9.18 version112Appendix D Recommended Reading113D.1 References for MPI113D.2 Books for Learning MPI Programming113D.3 Reference and Source for SLURM113D.4 InfiniBand113D.5 OpenFabrics113D.6 Clusters114D.7 Rocks114Appendix E Glossary115Index121文件大小: 688.7 KB页数: 122Language: English打开用户手册