/IPM

Integrated Performance Monitoring for High Performance Computing

Primary LanguageCGNU Lesser General Public License v2.1LGPL-2.1

IPM

Integrated Performance Monitoring for HPC

Quick start

IPM uses the common GNU autotools build system for compilation. As such, a typical install of IPM from this repository must begin by generating the autoconf configure script via our provided bootstrap.sh script:

./bootstrap.sh

Once this is done, you should be able to run

./configure
make
make install

As usual, you will likely want to examine key options available for the configure script, which can be viewed by running ./configure --help. For example, you can enable support for hardware performance counter collection via PAPI by specifying the PAPI installation path using the --with-papi=<path> option. You may also need to specify your MPI compiler wrappers via the MPICC and MPIFC variables - see ./configure --help for details.

Once you have built and installed IPM, using it to profile your application should be fairly straightforward. Suppose that $PREFIX is the installation prefix you used when installing IPM. In this case, you can enable profiling by appending

-L$PREFIX/lib -lipm

to your link line (making sure it is the last argument). If you are linking an application that uses the Fortran bindings to MPI, then you will also need to add -lipmf, making sure that it comes before -lipm:

-L$PREFIX/lib -lipmf -lipm

As usual, if you are using dynamic linking and $PREFIX/lib has not been added to your $LD_LIBRARY_PATH, then you will likely also want to pass the former to the linker as an -rpath.

Note: The master branch of IPM is in general usable, but should be considered development software. We strongly recommend that for production use you build a tagged release of IPM. The most recent release is 2.0.6.

Examples

Suppose that you are building a pure C MPI code contained in a single source file that you would like to profile with IPM, and that mpicc is the MPI C compiler wrapper for your system. In this case, you could simply run:

mpicc my_code.c -o my_code.x -L$PREFIX/lib -lipm

to produce an IPM instrumented executable.

Now suppose that you are building a mixed C and Fortran MPI code that you would like to profile with IPM, and that mpicc and mpifort are the MPI C and Fortran compiler wrappers for your system. In this case, you could use:

mpicc -c my_c_routines.c -o my_c_routines.o
mpifort -c my_fort_routines.f90 -o my_fort_routines.o
mpifort my_c_routines.o my_fort_routines.o -o my_code.x -L$PREFIX/lib -lipmf -lipm

About IPM

IPM is a portable profiling infrastructure for parallel codes. It provides a low-overhead profile of application performance and resource utilization in a parallel program. Communication, computation, and IO are the primary focus. While the design scope targets production computing in HPC centers, IPM has found use in application development, performance debugging and parallel computing education. The level of detail is selectable at runtime and presented through a variety of text and web reports.

IPM has extremely low overhead, is scalable and easy to use requiring no source code modification. It runs on Cray XT/XE, IBM Blue Gene, most Linux clusters using MPICH/OPENMPI, SGI Altix and some NEC machines. IPM is available under an Open Source software license (LGPL). It is currently installed on several Teragrid, Department of Energy, and other supercomputing resources.

IPM brings together several types of information important to developers and users of parallel HPC codes. The information is gathered in a way the tries to minimize the impact on the running code, maintaining a small fixed memory footprint and using minimal amounts of CPU. When the profile is generated the data from individual tasks is aggregated in a scalable way.

IPM is modular. You can measure just what you want. In addition to the core timings and benchmarking data, IPM currently delivers the following modules:

  • MPI: Communication topology and statistics for each MPI call and buffer size.
  • HPM: FLOPs and such via PAPI on-chip event counters.
  • OpenMP: thread level details about load imbalance.
  • Memory: memory high watermark, getrusage, sbrk.
  • **Switch:**Communication volume and packet loss.
  • File I/O: Data written to and read from disk.
  • GPU: Memory copied in/out of the GPU and time in kernels.
  • Power: Joules of energy consumed by the app.

The 'integrated' in IPM is multi-faceted. It refers to combining the above information together through a common interface and also the integration of the records from all the parallel tasks into a single report. At a high level we seek to integrate together the information useful to all stakeholders in HPC into a common interface that enables a deeper understanding. This includes application developers, science teams using applications, HPC managers, and system architects.

Known issues

The following are known issues affecting the current master branch of IPM:

Incorrect MPI message sizes in certain circumstances

  1. IPM may report incorrect message sizes for the family of MPI_Wait and MPI_Test functions. This is because IPM indiscriminately extracts the message size from the status object of every single MPI_Wait and MPI_Test call. However, there are at least 3 situations when the status object is either undefined or partially defined.

    • The request handle is from a non-blocking send

      "The fields in a status object returned by a call to MPI_WAIT, MPI_TEST, or any of the other derived functions (MPI_{TEST|WAIT}{ALL|SOME|ANY}), where the request corresponds to a send call, are undefined, with two exceptions: The error status field will contain valid information if the wait or test call returned with MPI_ERR_IN_STATUS ; and the returned status can be queried by the call MPI_TEST_CANCELLED." [MPI standard v3.1 page 52]

    • The request handle is from a non-blocking collective call

      "Upon returning from a completion call in which a nonblocking collective operation completes, the MPI_ERROR field in the associated status object is set appropriately, see Section 3.2.5. The values of the MPI_SOURCE and MPI_TAG fields are undefined." [MPI standard v3.1 page 197]

    • The request handle is from an RMA operation

      "Upon returning from a completion call in which an RMA operation completes, the MPI_ERROR field in the associated status object is set appropriately (see Section 3.2.5). All other fields of status and the results of status query functions (e.g., MPI_GET_COUNT) are undefined." [MPI standard v3.1 page 430]

    The MPI implementation may still provide the message size for these special cases but it is not required by the standard. For example, we have found that openmpi-1.8.1 initializes status data corresponding to send requests, while mpich-2.1.5 and mpich-3.3a2 do not.

  2. IPM will not collect the message size for an MPI function that is passed MPI_STATUS_IGNORE or MPI_STATUSES_IGNORE. This affects receive, probe, wait and test functions. The only receive function that is not affected by this issue is MPI_Irecv.

  3. IPM may report incorrect message sizes for MPI_Irecv because the message size is calculated from the receive count and datatype arguments. This may be larger than the eventual message because it is legitimate for the sender to send a shorter message.

    Key takeaway: IPM is only expected to collect correct message sizes for MPI send and MPI collective functions. The message size data for other MPI functions should be ignored unless you know that your application is not affected by one of the above issues.