10. Resilience

This section documents the features and options specific to the PRTE Level Fault Tolerant PMIx reference RunTime Environment (PRTE)

10.1. Features

This implementation provides a runtime level failure detection and propagation mechanism for both process and node failure.

10.1.1. What’s added to support fault-tolerance?

  1. New module under src/mca/errmgr: detector: Daemons monitor one another along a ring topology to detect node failures. src/mca/odls is in charge of detecting the failure of locally hosted processes (using SIGCHLD signals from the operating system).

  2. New component: propagate with module prperror: Prepares the content of the reliable broadcast messages (i.e., the list of failed processes). In order to populate the list of failed processes in node failure cases, the list of processes hosted by a particular daemon is collected by prperror module.

  3. New module under src/mca/grpcomm: bmg The BMG component implements a broadcast algorithm in a reliable way; to be noted, this component abides by the normal interface for a daemon broadcast and can reliably broadcast any type of information

  4. Test case for process failure under example/error_notify.c This test uses kill(pid) to kill a process to simulate process failure.

  5. Test case for daemon/node failure under example/daemon_error_notify.c This test uses kill(ppid) to kill a process’s parent to simulate node failure.

10.1.2. Building

./autogen.pl

# If you want to run mpi applications, you mpi and PRTE
# should have the same version of PMIx and libevent

./configure --enable-prte-ft --prefix=...  --with-pmix=/external-pmix-path --with-libevent=/external-libevent-path

make [-j N] all install
#    use an integer value of N for parallel builds

10.1.3. Running

10.1.3.1. Building your application

Compile your application as usual

  1. using the provided pcc for pmix-based application;

  2. using your mpicc for mpi-based application with a prte-based MPI (e.g., Open MPI).

10.1.3.2. Running your application

If running standalone:

  1. you need to launch first the DVM daemons with prte --mca prte_enable_ft true.

  2. You can then launch your application with by simply using the provided prun --enable-recovery.

Make sure to set your PATH and LD_LIBRARY_PATH properly.

If running with a PRRTE-based MPI (e.g., Open MPI):

  1. use mpiexec --enable-recovery --mca prte_enable_ft true.

10.1.3.3. Running under a batch scheduler

This code can operate under a job/batch scheduler, and is tested routinely with Slurm. One difficulty comes from the fact that many job schedulers will “cleanup” the application as soon as a process fails. In order to avoid this problem, it is preferred that you use -k, --no-kill [=off]: Do not automatically terminate a job if one of the nodes it has been allocated fails. within an allocation (e.g., salloc, sbatch) rather than a direct launch (e.g. srun).

10.1.4. Run-time tuning knobs

This code comes with a variety of knobs for controlling how it runs. The default parameters are sane and should result in very good performance in most cases. You can change those default by --prtemca parameter value:

  • prte_enable_recovery <true|false> (default: false) controls automatic cleanup of apps with failed processes.

  • prte_abort_non_zero_exit <true|false> (default: true) controls the job termination after a error occurred.

  • errmgr_detector_enable <true|false> (default: false) enable or disable error detection and propagation.

  • errmgr_detector_heartbeat_period <float> (default:5s) heartbeat period. Recommended value is 1/2 of the timeout.

  • errmgr_detector_heartbeat_timeout <float> (default:10s) heartbeat timeout (i.e. failure detection speed). Recommended value is 2 times the heartbeat period

To be noted: if you want to use prte failure detection and propagation features.

You MUST set prte_enable_recovery to true, prte_abort_non_zero_exit to false.

10.2. Testing

# Step 1
salloc -k -N num_of_nodes -w host1,host2...
     -k, --no-kill do not kill job on node failure

# Step 2
prte --prtemca prte_enable_ft true \
     --prtemca errmgr_detector_heartbeat_period 0.5  \
     --prtemca errmgr_detector_heartbeat_timeout 1  \
     --prtemca errmgr_detector_enable 1 \
     --prtemca prte_abort_on_non_zero_status 0 \
     --debug-daemons

# using 'errmgr_detector_enable 1' choose enable the error detector.

Config with --enable-debug, --debug-daemons will give you lots of information.

Also, the ring detector heartbeat sending frequency is not hard coded, you can change heartbeat_peroid and heartbeat_timeout by using MCA params. For example:

  • using --prtemca errmgr_detector_heartbeat_period 10 set the sending frequency to every 10 seconds(default is 5s)

  • using --prtemca errmgr_detector_heartbeat_timeout 20 set timeout to 20 seconds(default is 10s)

Step 3: under example we have 2 test codes error_notify.c, daemon_error_notify.c:

# Compile the codes
pcc -g error_notify.c -o error_notify

# Run
prun --oversubscribe --merge-stderr-to-stdout \
     --map-by node:DISPLAY:DISPLAYALLOC \
     --report-bindings --enable-recovery \
     --max-restarts 4 \
     --continuous -np num_of_procs error_notify -v

If use external pmix:

# Compile
pcc error_notify.c -o error_notify_1 \
     -I/external_pmix_install_path/include \
     -L/external_pmix_install_path/lib \
     -lpmix

# Run
prun --oversubscribe -x LD_LIBRARY_PATH \
     --merge-stderr-to-stdout \
     --map-by node:DISPLAY:DISPLAYALLOC \
     --report-bindings --enable-recovery \
     --max-restarts 4 \
     --continuous -np num_of_procs error_notify_1 -v

Iif use external pmix:

# Compile
pcc daemon_error_notify.c -o daemon_error_notify_1 \
     -I/external_pmix_install_path/include \
     -L/external_pmix_install_path/lib \
     -lpmix

# Run
prun --oversubscribe -x LD_LIBRARY_PATH \
     --merge-stderr-to-stdout \
     --map-by node:DISPLAY:DISPLAYALLOC \
     --report-bindings --enable-recovery \
     --max-restarts 4 \
     --continuous -np num_of_procs \
     daemon_error_notify_1 -v