Using the Hydra Process Manager

https://wiki.mpich.org/mpich/index.php/Using_the_Hydra_Process_Manager

Using the Hydra Process Manager

Jump to: navigation, search

This wiki page only provides information on the external usage of Hydra. If you are looking for the internal workings of Hydra, you can find it here.

General

Hydra is a process management system for starting parallel jobs. Hydra is designed to natively work with multiple daemons such as ssh, rsh, pbs, slurm and sge.

Quick Start

Once built, the Hydra executables are in the bin subdirectory of the install directory if you have done an install. You should put this (bin) directory in your PATH in your .cshrc or .bashrc for usage convenience:

Put in .cshrc:  setenv PATH /home/you/mpich/bin:$PATH

Put in .bashrc: export PATH=/home/you/mpich/bin:$PATH

To compile your application use mpicc:

shell$ mpicc app.c -o app

Create a file with the names of the machines that you want to run your job on. This file may or may not include the local machine.

shell$ cat hosts
   donner
   foo
   shakey
   terra

To run your application on these nodes, use mpiexec:

shell$ mpiexec -f hosts -n 4 ./app

If the same application is run with 8 processes, it will use the hosts in a round robin manner. For instance, with the above host file, ranks 0 and 4 will execute on donner, ranks 1 and 5 on foo, ranks 2 and 6 on shakey and ranks 3 and 7 on terra.

The host file can also be specified as follows:

shell$ cat hosts
   donner:2
   foo:2
   shakey:2
   terra:2

In this case, ranks 0-1 will execute on donner, ranks 2-3 on foo, ranks 4-5 on shakey and ranks 6-7 on terra.

A more complex host file layout can be:

shell$ cat hosts
   donner:2
   foo:3
   shakey:2

In this case, the first 2 processes are scheduled on "donner", thenext 3 on "foo" and the last 2 on "shakey". Comments in thehost file start with a "#" character.

shell$ cat hosts
   # This is a sample host file
   donner:2     # The first 2 procs are scheduled to run here
   foo:3        # The next 3 procs run on this host
   shakey:2     # The last 2 procs run on this host

Environment Settings

By default, hydra passes inherits environment variables from the shell on which mpiexec is launched, and passes it to the launched processes. However, each launcher disables the automatic propagation of some environment variables. For example, the SSH launcher disables the propagation of "DISPLAY", the SLURM launcher disables the propagation of "SLURM_*" variables, etc.

Users can force an environment variable to be propagated using the -genv option.

Environment variables can also be used to control several settings in Hydra:

HYDRA_HOST_FILE: This variable points to the default host file to use, when the "-f" option is not provided to mpiexec.

  For bash:
    export HYDRA_HOST_FILE=<path_to_host_file>/hosts

  For csh/tcsh:
    setenv HYDRA_HOST_FILE <path_to_host_file>/hosts

HYDRA_DEBUG: Setting this to "1" enables debug mode; set it to "0" to disable.

HYDRA_ENV: Setting this to "all" will pass all of the launching node's environment to the application processes. By default, if nothing is set, the launching node's environment is passed to the executables, as long as it does not overwrite any of the environment variables that have been preset by the remote shell.

HYDRA_LAUNCH_EXTRA_ARGS: The value of this environment variable is appended to the launcher. This can be used, for example, to pass the "-q" argument to the launcher by setting HYDRA_LAUNCH_EXTRA_ARGS=-q.

MPIEXEC_TIMEOUT: The value of this environment variable is the maximum number of seconds this job will be permitted to run. When time is up, the job is aborted.

MPIEXEC_PORT_RANGE: If this environment variable is defined then Hydra will restrict its usage of ports for connecting its various processes to ports in this range. If this variable is not assigned, but MPICH_PORT_RANGE is assigned, then it will use the range specified by MPICH_PORT_RANGE for its ports. Otherwise, it will use whatever ports are assigned to it by the system. Port ranges are given as a pair of integers separated by a colon.

HYDRA_PROXY_RETRY_COUNT: The value of this environment variable determines the number of retries a proxy does to establish a connection to the main server.

Hydra with Non-Ethernet Networks

There are two ways to use Hydra with TCP/IP over the non-default network.

The first way is using the -iface option to mpiexec to specify which network interface to use. For example, if your Myrinet network's IP emulation is configured on myri0, you can use:

shell$ mpiexec -f hostfile -iface myri0 -n 4 ./app1

Similarly, if your InfiniBand network's IP emulation is configured on ib0, you can use:

shell$ mpiexec -f hostfile -iface ib0 -n 4 ./app1

You can also control this using the HYDRA_IFACE environment variable.

The second way is to specify the appropriate IP addresses in your hostfile.

shell$ /sbin/ifconfig

eth0      Link encap:Ethernet  HWaddr 00:14:5E:57:C4:FA  
          inet addr:192.148.3.182  Bcast:192.148.248.255  Mask:255.255.255.0
          inet6 addr: fe80::214:5eff:fe57:c4fa/64 Scope:Link
          UP BROADCAST RUNNING MULTICAST  MTU:1500  Metric:1
          RX packets:989925894 errors:0 dropped:7186 overruns:0 frame:0
          TX packets:1480277023 errors:0 dropped:0 overruns:0 carrier:0
          collisions:0 txqueuelen:1000 
          RX bytes:441568994866 (411.2 GiB)  TX bytes:1864173370054 (1.6 TiB)
          Interrupt:185 Memory:e2000000-e2012100 

myri0     Link encap:Ethernet  HWaddr 00:14:5E:57:C4:F8  
          inet addr:10.21.3.182  Bcast:10.21.255.255  Mask:255.255.0.0
          inet6 addr: fe80::214:5eff:fe57:c4f8/64 Scope:Link
          UP BROADCAST RUNNING MULTICAST  MTU:1500  Metric:1
          RX packets:3068986439 errors:0 dropped:7841 overruns:0 frame:0
          TX packets:2288060450 errors:0 dropped:0 overruns:0 carrier:0
          collisions:0 txqueuelen:1000 
          RX bytes:3598751494497 (3.2 TiB)  TX bytes:1744058613150 (1.5 TiB)
          Interrupt:185 Memory:e4000000-e4012100 

ib0       Link encap:Ethernet  HWaddr 00:14:5E:57:C4:F8  
          inet addr:20.31.3.182  Bcast:10.21.255.255  Mask:255.255.0.0
          inet6 addr: fe80::214:5eff:fe57:c4f8/64 Scope:Link
          UP BROADCAST RUNNING MULTICAST  MTU:1500  Metric:1
          RX packets:3068986439 errors:0 dropped:7841 overruns:0 frame:0
          TX packets:2288060450 errors:0 dropped:0 overruns:0 carrier:0
          collisions:0 txqueuelen:1000 
          RX bytes:3598751494497 (3.2 TiB)  TX bytes:1744058613150 (1.5 TiB)
          Interrupt:185 Memory:e4000000-e4012100 

In the above case the 192.148.x.x IP series refers to the standard Ethernet (or Gigabit Ethernet) network, the 10.21.x.x series refers to Myrinet and the 10.31.x.x refers to InfiniBand.

shell$ cat hostfile-eth
   192.148.3.180
   192.148.3.181
   192.148.3.182
   192.148.3.183

shell$ cat hostfile-myri
   10.21.3.180
   10.21.3.181
   10.21.3.182
   10.21.3.183

shell$ cat hostfile-ib
   10.31.3.180
   10.31.3.181
   10.31.3.182
   10.31.3.183

To run over the Ethernet interface use:

shell$ mpiexec -f hostfile-eth -n 4 ./app1

To run over the Myrinet interface use:

shell$ mpiexec -f hostfile-myri -n 4 ./app1

Resource Managers and Launchers

Hydra natively interacts with a number of resource managers and launchers.

Resource managers provide information about the resources allocated by the user (e.g., slurm, pbs, loadleveler, lsf, sge, cobalt). Launchers allow mpiexec to launch processes on the system (e.g., ssh, rsh, fork, slurm, pbs, loadleveler, lsf, sge). Some tools act as both resource managers and launchers, while others play just one role.

In most cases, you don't need to do anything special for Hydra to work with your slurm, pbs, loadleveler, lsf, sge or cobalt environment. It'll automatically detect these resource managers and interact with them. The following command will be sufficient to launch processes on all the allocated cores:

shell$ mpiexec ./app

The above command will allow Hydra to inherit information about the resource manager and launcher available, resources allocated from the resource manager, and finally launch the appropriate number of processes.

However, if you do not want Hydra to auto-detect the launcher, you can customize it using the -launcher flag to mpiexec or the HYDRA_LAUNCHER environment variable.

shell$ mpiexec -launcher ssh -f hosts -n 4 ./app

 (or)

shell$ mpiexec -launcher fork -f hosts -n 4 ./app

The executable to use as the launcher can also be customized using the option -launcher-exec or the HYDRA_LAUNCHER_EXEC environment variable:

 $ mpiexec -launcher ssh -launcher-exec /usr/bin/ssh -f hosts -n 4 ./app

Similarly, if you do not want Hydra to auto-detect the resource manager, you can customize using the -rmk flag to mpiexec or the HYDRA_RMK environment variable.

shell$ mpiexec -rmk pbs ./app

Finally, if you do not want Hydra to auto-detect the number of processes allocated or the machines allocated, you can customize that using the -n and -f flags.

shell$ mpiexec -rmk pbs -n 4 -f ~/hosts ./app

Process-core Binding

On supported platforms, Hydra automatically configures available process-core binding capability (currently using PLPA or hwloc). We support multiple levels of allocation strategies:

Basic allocation strategies: There are two forms of basic allocation: (i) based on a round-robin mechanism using the OS specified processor IDs, and (ii) based on a user-defined mapping. Further, for the user-defined mapping, two schemes are provided---command-line and host-file based. The command-line scheme lets the user specify a common-mapping for all physical nodes on the command line. The host-file scheme is the most general and lets the user specify the mapping for each node separately.

The modes of process binding in the basic allocation are: round-robin ("rr") and user-defined ("user").

shell$ mpiexec -binding rr -f hosts -n 8 ./app

Within the user-defined binding, two modes are supported: command-line and host-file based. The command-line based mode can be used as follows:

shell$ mpiexec -binding user:0,3 -f hosts -n 4 ./app

If a machine has 4 processing elements, and only two bindings are provided (as in the above example), the rest are padded with (-1), which refers to no binding. Also, the mapping is the same for all machines; so if the application is run with 8 processes, the first 2 processes on "each machine" are bound to processing elements as specified.

The host-file based mode for user-defined binding can be used by the "binding=" argument on each host line. E.g.:

shell$ cat hosts
   donner:4    binding=user:0,-1,-1,3
   foo:4       binding=rr
   shakey:2

Using this method, each host can be given a different mapping. Any unspecified mappings are treated as (-1), referring to no binding.

Command-line based mappings are given a higher priority than the host-file based mappings. So, if a mapping is given at both places, the host-file mappings are ignored.

Topology-aware allocation strategies: These are a bit more intelligent in that they try to understand the system processing unit topology and assign processes in that order.

Different modes of process binding in the topology-aware allocation are:

CPU based options:

1. sockets -- allocate processes to one socket at a time (allocating all processing units on the socket to each process)
2. cores -- allocate processes to one core at a time (allocating all processing units on the core to each process)
3. rr:sockets -- allocate processes to one socket at a time (rotating between the processing units on each socket)
4. rr:cores -- allocate processes to one core at a time (rotating between the processing units on each core)

Cache based options:

1. l1 -- allocate processes to one L1 cache domain at a time (allocating all processing units on the cache domain to each process)
2. l2 -- allocate processes to one L2 cache domain at a time (allocating all processing units on the cache domain to each process)
3. l3 -- allocate processes to one L3 cache domain at a time (allocating all processing units on the cache domain to each process)
4. rr:l1 -- allocate processes to one L1 cache domain at a time (rotating between the processing units on each cache domain)
5. rr:l2 -- allocate processes to one L2 cache domain at a time (rotating between the processing units on each cache domain)
6. rr:l3 -- allocate processes to one L3 cache domain at a time (rotating between the processing units on each cache domain)

shell$ mpiexec -binding sockets -f hosts -n 8 ./app

(or)

shell$ mpiexec -binding cores -f hosts -n 8 ./app

(or)

shell$ mpiexec -binding l2 -f hosts -n 6 ./app

(or)

shell$ mpiexec -binding rr:l1 -f hosts -n 6 ./app

Consider the following layout of processing elements in the system (e.g., two nodes, each with two processors, and each processor withtwo cores). Suppose the Operating System assigned processor IDs for each of these processing elements are as shown below:

__________________________________________      __________________________________________
|  _________________    _________________  |    |  _________________    _________________  | 
| |  _____   _____  |  |  _____   _____  | |    | |  _____   _____  |  |  _____   _____  | |
| | |     | |     | |  | |     | |     | | |    | | |     | |     | |  | |     | |     | | |
| | |     | |     | |  | |     | |     | | |    | | |     | |     | |  | |     | |     | | | 
| | |  0  | |  2  | |  | |  1  | |  3  | | |    | | |  0  | |  2  | |  | |  1  | |  3  | | |
| | |     | |     | |  | |     | |     | | |    | | |     | |     | |  | |     | |     | | |
| | |_____| |_____| |  | |_____| |_____| | |    | | |_____| |_____| |  | |_____| |_____| | |
| |_________________|  |_________________| |    | |_________________|  |_________________| |
|__________________________________________|    |__________________________________________|

In this case, the binding options are as follows:

1. rr 0, 1, 2, 3 (use the order provided by the OS)
2. sockets 0+2, 1+3 (allocate all cores on the socket to each process)
3. cores 0, 2, 1, 3 (allocate all processing units on a core to each process: a core has just one processing unit in this case)
4. rr:sockets 0, 1, 2, 3 (allocate one core on the socket at a time to each process)
5. rr:cores 0, 2, 1, 3 (allocate one processing unit on a core at a time to each process: a core has just one processing unit in this case)
6. user as defined by the user

Binding options can also be controlled with the environment variable HYDRA_BINDING. The binding library to use can be controlled with the -bindlib argument to mpiexec or the HYDRA_BINDLIB environment variable.

X Forwarding

X-forwarding is specific to each bootstrap server. Some servers do it by default, while some don't. For ssh, this is disabled by default. To force-enable it, you should use the option -enable-x to mpiexec.

shell$ mpiexec -enable-x -f hosts -n 4 ./app

Checkpoint/Restart Support

Hydra provides checkpoint/restart capability. Currently, only BLCR is supported. To use checkpointing include the -ckpointlib option for mpiexec to specify the checkpointing library to use and -ckpoint-prefix to specify the directory where the checkpoint images should be written:

shell$ mpiexec -ckpointlib blcr -ckpoint-prefix /home/buntinas/ckpts/app.ckpoint -f hosts -n 4 ./app

While the application is running, the user can request for a checkpoint at any time by sending a SIGUSR1 signal to mpiexec.

You can also automatically checkpoint the application at regular intervals using the mpiexec option -ckpoint-interval to specify the number of seconds between checkpoints:

shell$ mpiexec -ckpointlib blcr -ckpoint-prefix /home/buntinas/ckpts/app.ckpoint -ckpoint-interval 3600 -f hosts -n 4 ./app

The checkpoint/restart parameters can be controlled with the environment variables HYDRA_CKPOINTLIB, HYDRA_CKPOINT_PREFIX and HYDRA_CKPOINT_INTERVAL.

Each checkpoint generates one file per node. Note that checkpoints for all processes on a node will be stored in the same file. Each time a new checkpoint is taken an additional set of files are created. The files are numbered by the checkpoint number. This allows the application to be restarted from checkpoints other than the most recent. The checkpoint number can be specified with the -ckpoint-num parameter. To restart a process:

shell$ mpiexec -ckpointlib blcr -ckpoint-prefix /home/buntinas/ckpts/app.ckpoint -ckpoint-num 5 -f hosts -n 4

Note that by default, the process will be restarted from the first checkpoint, so in most cases, the checkpoint number should be specified.

Demux Engines

Hydra supports multiple I/O demux engines including poll and select. The default is "poll". You can pick these through the mpiexec option -demux:

shell$ mpiexec -demux select -f hosts -n 4 ./app

This can also be controlled by using the HYDRA_DEMUX environment variable.

Debugging

Hydra natively supports parallel debuggers such as totalview and DDT.

You can debug the MPI application with totalview by launching it as:

shell$ totalview mpiexec -a -f hosts -n 4 ./app

The "-a" option is a totalview parameter which tells it that the arguments after that need to be passed to mpiexec.

You can debug the MPI application with DDT by launching it as:

shell$ ddt mpiexec -a -f hosts -n 4 ./app

The "-a" option is a DDT parameter which tells it that the arguments after that need to be passed to mpiexec.

For serial debuggers such as 'ddd', you can launch each process with a separate debugger window as:

shell$ mpiexec -f hosts -n 4 ddd ./app

This will spawn four copies of "./app" each one launched under a separate 'ddd' instance.

You can do something similar to 'ddd' for serial debuggers that do not have a graphical interface, such as 'gdb', by launching them over 'xterm':

shell$ mpiexec -f hosts -n 4 xterm -e gdb ./app

If you want only the third process to run in a debugger, and the remaining processes to run normally, you can do:

shell$ mpiexec -f hosts -n 2 ./app : -n 1 xterm -e gdb ./app : -n 1 ./app

Using Hydra on Machines with Different User Names

Hydra only supports using different usernames on some launchers (such as ssh and rsh). For this, the host file should contain a "user=<foo>" entry.

shell$ cat hosts
   donner  user=foo
   foo     user=bar
   shakey  user=bar
   terra   user=foo