13.8. General Tuning

Error

This section represents old content from the <= v4.x FAQ that has not been properly converted to the new-style documentation. The content here was perfunctorily converted to RST, but it still needs to be:

  1. Converted from a question-and-answer style to a regular documentation style (like the rest of these docs).

  2. Removed from this section and folded into other sections in these docs.

To be clear, this section will eventually be deleted; do not write any new content in this section.

13.8.1. How do I install my own components into an Open MPI installation?

By default, Open MPI looks in two places for components at run-time (in order):

  1. $prefix/lib/openmpi/: This is the system-provided components directory, part of the installation tree of Open MPI itself.

  2. $HOME/.openmpi/components/: This is where users can drop their own components that will automatically be “seen” by Open MPI at run-time. This is ideal for developmental, private, or otherwise unstable components.

Note that the directories and search ordering used for finding components in Open MPI is, itself, an MCA parameter. Setting the mca_component_path changes this value (a colon-delimited list of directories).

Note also that components are only used on nodes where they are “visible”. Hence, if your $prefix/lib/openmpi/ is a directory on a local disk that is not shared via a network filesystem to other nodes where you run MPI jobs, then components that are installed to that directory will only be used by MPI jobs running on the local node.

More specifically: components have the same visibility as normal files. If you need a component to be available to all nodes where you run MPI jobs, then you need to ensure that it is visible on all nodes (typically either by installing it on all nodes for non-networked filesystem installs, or by installing them in a directory that is visible to all nodes via a networked filesystem). Open MPI does not automatically send components to remote nodes when MPI jobs are run.

13.8.2. What is processor affinity? Does Open MPI support it?

Open MPI supports processor affinity on a variety of systems through process binding, in which each MPI process, along with its threads, is “bound” to a specific subset of processing resources (cores, packages, etc.). That is, the operating system will constrain that process to run on only that subset.

Note

The operating system may allow other processes to run on the same resources.

Affinity can improve performance by inhibiting excessive process movement — for example, away from “hot” caches or NUMA memory. Judicious bindings can improve performance by reducing resource contention (by spreading processes apart from one another) or improving interprocess communications (by placing processes close to one another). Binding can also improve performance reproducibility by eliminating variable process placement.

Warning

Processor affinity probably should not be used when a node is over-subscribed (i.e., more processes are launched than there are processors).

This can lead to a serious degradation in performance (even more than simply oversubscribing the node). Open MPI will usually detect this situation and automatically disable the use of processor affinity (and display run-time warnings to this effect).

13.8.3. What is memory affinity? Does Open MPI support it?

Memory affinity is increasingly relevant on modern servers because most architectures exhibit Non-Uniform Memory Access (NUMA) architectures. In a NUMA architecture, memory is physically distributed throughout the machine even though it is virtually treated as a single address space. That is, memory may be physically local to one or more processors — and therefore remote to other processors.

Simply put: some memory will be faster to access (for a given process) than others.

Open MPI supports general and specific memory affinity, meaning that it generally tries to allocate all memory local to the processor that asked for it. When shared memory is used for communication, Open MPI uses memory affinity to make certain pages local to specific processes in order to minimize memory network/bus traffic.

Open MPI supports memory affinity on a variety of systems.

In recent versions of Open MPI, memory affinity is controlled through the Hardware Locality (hwloc) library.

Note that memory affinity support is enabled only when processor affinity is enabled. Specifically: using memory affinity does not make sense if processor affinity is not enabled because processes may allocate local memory and then move to a different processor, potentially remote from the memory that it just allocated.

13.8.4. How do I tell Open MPI to use processor and/or memory affinity?

Open MPI will, by default, enable processor and memory affinity when not running in an oversubscribed environment (i.e., when the number of MPI processes are less than or equal two the number of processors available).

The mpirun(1) man page for each version of Open MPI contains a lot of information about the use of processor and memory affinity. You should consult the mpirun(1) page for your version of Open MPI for detailed information about processor/memory affinity.

Error

TODO Link to mpirun(1) …?

13.8.5. Does Open MPI support calling fork(), system(), or popen() in MPI processes?

It depends on a lot of factors, including (but not limited to):

  • The operating system

  • The underlying compute hardware

  • The network stack

  • Interactions with other middleware in the MPI process

In some cases, Open MPI will determine that it is not safe to fork(). In these cases, Open MPI will register a pthread_atfork() callback to print a warning when the process forks.

This warning is helpful for legacy MPI applications where the current maintainers are unaware that system() or popen() is being invoked from an obscure subroutine nestled deep in millions of lines of Fortran code (we’ve seen this kind of scenario many times).

However, this atfork handler can be dangerous because there is no way to unregister an atfork handler. Hence, packages that dynamically open Open MPI’s libraries (e.g., Python bindings for Open MPI) may fail if they finalize and unload libmpi, but later call fork. The atfork system will try to invoke Open MPI’s atfork handler; nothing good can come of that.

For such scenarios, or if you simply want to disable printing the warning, Open MPI can be set to never register the atfork handler with the mpi_warn_on_fork MCA parameter. For example:

shell$ mpirun --mca mpi_warn_on_fork 0 ...

Of course, systems that dlopen("libmpi.so", ...) may not use Open MPI’s mpirun, and therefore may need to use (JMS: this ref no longer exists – it moved to elsewhere) a different mechanism to set MCA parameters <faq-general-tuning-setting-mca-params>`.

13.8.6. I want to run some performance benchmarks with Open MPI. How do I do that?

Running benchmarks is an extremely difficult task to do correctly. There are many, many factors to take into account; it is not as simple as just compiling and running a stock benchmark application. This documentation is by no means a definitive guide, but it does try to offer some suggestions for generating accurate, meaningful benchmarks.

  1. Decide exactly what you are benchmarking and setup your system accordingly. For example, if you are trying to benchmark maximum performance, then many of the suggestions listed below are extremely relevant (be the only user on the systems and network in question, be the only software running, use processor affinity, etc.). If you’re trying to benchmark average performance, some of the suggestions below may be less relevant. Regardless, it is critical to know exactly what you’re trying to benchmark, and know (not guess) both your system and the benchmark application itself well enough to understand what the results mean.

    To be specific, many benchmark applications are not well understood for exactly what they are testing. There have been many cases where users run a given benchmark application and wrongfully conclude that their system’s performance is bad — solely on the basis of a single benchmark that they did not understand. Read the documentation of the benchmark carefully, and possibly even look into the code itself to see exactly what it is testing.

    Case in point: not all ping-pong benchmarks are created equal. Most users assume that a ping-pong benchmark is a ping-pong benchmark is a ping-pong benchmark. But this is not true; the common ping-pong benchmarks tend to test subtly different things (e.g., NetPIPE, TCP bench, IMB, OSU, etc.). Make sure you understand what your benchmark is actually testing.

  2. Make sure that you are the only user on the systems where you are running the benchmark to eliminate contention from other processes.

  3. Make sure that you are the only user on the entire network / interconnect to eliminate network traffic contention from other processes. This is usually somewhat difficult to do, especially in larger, shared systems. But your most accurate, repeatable results will be achieved when you are the only user on the entire network.

  4. Disable all services and daemons that are not being used. Even “harmless” daemons consume system resources (such as RAM) and cause “jitter” by occasionally waking up, consuming CPU cycles, reading or writing to disk, etc. The optimum benchmark system has an absolute minimum number of system services running.

  5. Ensure that processor and memory affinity are properly utilized to disallow the operating system from swapping MPI processes between processors (and causing unnecessary cache thrashing, for example).

    Warning

    On NUMA architectures, having the processes getting bumped from one socket to another is more expensive in terms of cache locality (with all of the cache coherency overhead that comes with the lack of it) than in terms of memory transfer routing (see below).

  6. Be sure to understand your system’s architecture, particularly with respect to the memory, disk, and network characteristics, and test accordingly. For example, on NUMA architectures, memory accesses may be routed through a memory interconnect; remote device and/or memory accesses will be noticeably slower than local device and/or memory accesses.

  7. Compile your benchmark with the appropriate compiler optimization flags. With some MPI implementations, the compiler wrappers (like mpicc, mpifort, etc.) add optimization flags automatically. Open MPI does not. Add -O or other flags explicitly.

  8. Make sure your benchmark runs for a sufficient amount of time. Short-running benchmarks are generally less accurate because they take fewer samples; longer-running jobs tend to take more samples.

  9. If your benchmark is trying to benchmark extremely short events (such as the time required for a single ping-pong of messages):

    • Perform some “warmup” events first. Many MPI implementations (including Open MPI) — and other subsystems upon which the MPI uses — may use “lazy” semantics to setup and maintain streams of communications. Hence, the first event (or first few events) may well take significantly longer than subsequent events.

    • Use a high-resolution timer if possible — gettimeofday() only returns millisecond precision (sometimes on the order of several microseconds).

    • Run the event many, many times (hundreds or thousands, depending on the event and the time it takes). Not only does this provide more samples, it may also be necessary, especially when the precision of the timer you’re using may be several orders of magnitude less precise than the event you’re trying to benchmark.

  10. Decide whether you are reporting minimum, average, or maximum numbers, and have good reasons why.

  11. Accurately label and report all results. Reproducibility is a major goal of benchmarking; benchmark results are effectively useless if they are not precisely labeled as to exactly what they are reporting. Keep a log and detailed notes about the ‘’exact’’ system configuration that you are benchmarking. Note, for example, all hardware and software characteristics (to include hardware, firmware, and software versions as appropriate).

13.8.7. I am getting a MPI_WIN_FREE error from IMB-EXT — what do I do?

When you run IMB-EXT with Open MPI, you’ll see a message like this:

[node01.example.com:2228] *** An error occurred in MPI_Win_free
[node01.example.com:2228] *** on win
[node01.example.com:2228] *** MPI_ERR_RMA_SYNC: error while executing rma sync
[node01.example.com:2228] *** MPI_ERRORS_ARE_FATAL (your MPI job will now abort)

This is due to a bug in the Intel MPI Benchmarks, known to be in at least versions v3.1 and v3.2. Intel was notified of this bug in May of 2009. If you have a version after then, the bug should be fixed. If not, here is the fix that you can apply to the IMB-EXT source code yourself.

Here is a small patch that fixes the bug in IMB v3.2:

diff -u imb-3.2-orig/src/IMB_window.c imb-3.2-fixed/src/IMB_window.c
--- imb-3.2-orig/src/IMB_window.c     2008-10-21 04:17:31.000000000 -0400
+++ imb-3.2-fixed/src/IMB_window.c      2009-07-20 09:02:45.000000000 -0400
@@ -140,6 +140,9 @@
                          c_info->rank, 0, 1, c_info->r_data_type,
                          c_info->WIN);
           MPI_ERRHAND(ierr);
           }
+          /* Added a call to MPI_WIN_FENCE, per MPI-2.1 11.2.1 */
+          ierr = MPI_Win_fence(0, c_info->WIN);
+          MPI_ERRHAND(ierr);
           ierr = MPI_Win_free(&c_info->WIN);
           MPI_ERRHAND(ierr);
           }

And here is the corresponding patch for IMB v3.1:

Index: IMB_3.1/src/IMB_window.c
===================================================================
--- IMB_3.1/src/IMB_window.c(revision 1641)
+++ IMB_3.1/src/IMB_window.c(revision 1642)
@@ -140,6 +140,10 @@
                          c_info->rank, 0, 1, c_info->r_data_type, c_info->WIN);
           MPI_ERRHAND(ierr);
           }
+          /* Added a call to MPI_WIN_FENCE here, per MPI-2.1
+             11.2.1 */
+          ierr = MPI_Win_fence(0, c_info->WIN);
+          MPI_ERRHAND(ierr);
           ierr = MPI_Win_free(&c_info->WIN);
           MPI_ERRHAND(ierr);
 }