-
Notifications
You must be signed in to change notification settings - Fork 868
HostFilePlan
As Ralph has pointed out, there is some confusion surrounding the use of ''--host'' and ''--hostfile'' within resource managed environments. This problem arises because of overloading what these parameters mean when combined with each other and with a resource managed environment.
In some cases ''--host'' and ''--hostfile'' define the resource list that the job should run on. In other cases, ''--host'' and ''--hostfile'' define a filter to the resource list. Note we are using the term filter because in certain scenarios, ''--host'' and ''--hostfile'' can be used to specify (i.e. filter) a subset of nodes to use in the resource list.
While the usage of ''--host'' is well defined in the [http://www.open-mpi.org/faq/?category=running#mpirun-host faq], the use of ''--hostfile'' is not. We therefore extend the meaning of how ''--hostfile'' should work in a resource managed environment. We also go further to cover some of the issues called out in ticket trac ticket 1018.
In this section, we define how a resource manager (RM), ''--host'', and ''--hostfile'' work together.
This table shows how things work in the absence of a RM.
How host and hostfile behave without RM:
--host |
--hostfile |
Result |
|---|---|---|
| unset | unset | job is run on mpirun node with np=1 |
| set | unset | host defines resource list |
| unset | set | hostfile defines resource list |
| set | set | hostfile defines resource list, host defines filter |
This table shows how things should work when running within a RM
How host and hostfile behave with RM:
--host |
--hostfile |
Result |
|---|
|
| unset | unset | job is run on all resources defined by RM | | set | unset | host defines filter to RM resource list | | unset | set | hostfile defines filter to RM resource list | | set | set | host and hostfile define filter to RM resource list |
To reiterate the points called out by the above tables, we walk through eight cases.
Case 1: mpirun foo
With nothing specified, one copy of foo is run on the same node as mpirun.
Case 2: mpirun --host foo
In this case, --host defines the host list. This host list is all
the nodes that the job foo will run on. Note that the job may not
run on all the nodes if -np is something smaller than the number
of nodes listed.
Case 3: mpirun --hostfile bar
In this case, --hostfile defines the host list.
Case 4: mpirun --host foo --hostfile bar
In this case, --hostfile defines the host list. The --host now
becomes a filter and specifies the subset of nodes that this job
will run on.
Imagine a shell was started as shown in the SGE example below.
qrsh -pe cre
Then, we can walk through the same four cases again but this time within the context of a RM.
Case 1: mpirun foo
With nothing specified, foo is run on all the resources defined by the RM.
Case 2: mpirun --host foo
In this case, --host defines a filter to the resource list defined by the
RM. The job will only run on those hosts.
Case 3: mpirun --hostfile bar
In this case, --hostfile defines a filter to the resource list defined by the
RM. The job will only run on those hosts.
Case 4: mpirun --host foo --hostfile bar
In this case, both --host and --hostfile define a filter and
specifies the subset of nodes that this job will run on.
In all the cases above, the filters must represent a subset of the resource list. Otherwise, one will get an error.
There have been requests to extend the initial available resource list in multi-app context and MPI_Comm_spawn cases. This is supposed to allow the user to specify a set of nodes outside an RM's control and thus not initially given as a resource list. One example is an mpirun specifying multiple app contexts in which one context is suppose to launch on a specific set of nodes not in the resource list. Another example, a call to MPI_Comm_spawn may want to spawn on a node that was not listed in the initial allocation. Currently (trunk as of r15584), this feature is not supported. However, as outlined in ticket trac ticket 1018, we propose adding two new mpirun parameters, --add-host and --add-hostfile. The definition of these parameters are defined in the ticket and suggested to be acceptable MPI_Info Keys to MPI_Comm_spawn{_multiple} below.
In all cases, the initial resource list will define where subsequent MPI_Comm_spawn jobs can run. The filters given to mpirun will only apply to the processes being launched. This also implies that in the unmanaged case, the initial call to mpirun must list all nodes that the job may run on. Because the resource list is inherited from the mpirun job both ''--host'' and ''--hostfile'' are always interpreted as filters when their MPI_Info Keys are passed in to MPI_Comm_spawn{multiple}.
Any mpirun that is executed under an existing universe behaves as if it was a MPI_Comm_spawn. This means that ''--host'' and ''--hostfile'' act as filters. One would have to use ''--add-host'' or ''--add-hostfile'' to expand the universe.
More needs to be analyzed here in the case of a hostfile that contains specific rank mappings.
In the following discussion, connected and disconnected refer to the definitions from the Release Connections chapter of the MPI_2 specification.
If two jobs are connected, and one calls MPI_Finalize and exit, then the behavior of both jobs is undefined. They may hang or abort.
If two jobs are disconnected, then the either job should be able to call MPI_Finalize and exit without affecting the other.
To support all of this in calls to MPI_Comm_spawn and MPI_Comm_spawn_multiple,
five new MPI_Info keys are proposed. The first four are analogous to what
is available to mpirun. The fifth key tells the child job to map itself
in the same way the parent is mapped.
host
hostfile
add-host
add-hostfile
parent-mapping
The host and hostfile keys will always act as filters and the add-host and add-hostfile
behave as if they were handed to a call to mpirun.
When being used as a filter, the ''--host'' switch will support the ''!^'' modifier which says that the filter will include all the nodes except this one. It will behave in much the same way as the MPI library supports it with btls. This means the ''!^'' applies to all the nodes in the list.
mpirun --host !^foo,bar,baz a.out says the job should run on all nodes except boo, bar, and baz.
mpirun --host !^fee,fi --hostfile hf a.out says the job should run on all the nodes listed
in the hostfile hf except for fee and fi.
Now, we are going to try and list all the different examples and what may happen. For the following examples, we have 3 nodes, each with 4 processors per node.
ct-0 slots=4
ct-1 slots=4
ct-2 slots=4
Also, imagine a hostfile with these entries
###################
# Hostfile ct
ct-0 slots=4
ct-1 slots=4
###################
> mpirun -hostfile ct /bin/hostname
ct-0
ct-0
ct-0
ct-0
ct-1
ct-1
ct-1
ct-1
> mpirun -host ct-1,ct-2 /bin/hostname
ct-1
ct-2
> mpirun -hostfile ct -host ct-1
ct-1
> mpirun -hostfile ct -host ct-2 /bin/hostname
--------------------------------------------------------------------------
Some of the requested hosts are not included in the current allocation for the
application:
/bin/hostname
The requested hosts were:
ct-2
Verify that you have mapped the allocated resources properly using the
--host specification.
--------------------------------------------------------------------------
> mpirun -hostfile ct -add-host ct2 /bin/hostname
ct-0
ct-0
ct-0
ct-0
ct-1
ct-1
ct-1
ct-1
ct-2
Imagine, we have a program that calls MPI_Comm_spawn, and that it uses the info keys to direct the spawning. For example,
MPI_Info_set(info, "host", "ct-1");
> mpirun -hostfile ct spawn-program
The spawn will use those keys in the same way a call to mpirun would have. So, in this case, the spawn would cause the MPI processes to run on ct-1. If the "host" info key was set to a host that was not in the initial allocation, then we would get an error.
In the following case, we show a multiple application example. As we can see, we are allowed to specify multiple hostfiles in this case. The hostfiles are therefore used as defined in the table at the beginning of this wiki.
mpirun -hostfile hf1 a.out1 : -hostfile hf2 a.out2 : -hostfile hf3 a.out3
> qstat -f
queuename qtype used/tot. load_avg arch states
----------------------------------------------------------------------------
all.q@ct-0 BIP 0/4 8.64 sol-sparc64
----------------------------------------------------------------------------
all.q@ct-1 BIP 0/4 1.09 sol-sparc64
> qsh -pe cre 8
waiting for interactive job to be scheduled ...
Your interactive job 279 has been successfully scheduled.
> mpirun hostname
ct-1
ct-1
ct-1
ct-1
ct-0
ct-0
ct-0
ct-0
> mpirun -np 6 hostname
ct-1
ct-1
ct-1
ct-1
ct-0
ct-0
> mpirun -host ct-0 hostname
ct-0
ct-0
ct-0
ct-0
> mpirun -host ct-1 hostname
ct-1
ct-1
ct-1
ct-1
> more /home/rolfv/hf.0
ct-0 slots=2
> mpirun -hostfile /home/rolfv/hf.0 hostname
ct-0
ct-0
burl-ct-v440-1 52 =>mpirun -host ct-2 hostname
--------------------------------------------------------------------------
Some of the requested hosts are not included in the current allocation for the
application:
hostname
The requested hosts were:
ct-2
Verify that you have mapped the allocated resources properly using the
--host specification.
--------------------------------------------------------------------------
In the following case, the behavior is resource manager dependent. Some resource managers may allow one to add a host to the list of available nodes whereas others may not. Here is an example of this not working under SGE.
burl-ct-v440-1 51 =>mpirun -add-host ct-2 hostname
error: executing task of job 279 failed:
[burl-ct-v440-1:24610] ERROR: A daemon on node ct-2 failed to start as expected.
[burl-ct-v440-1:24610] ERROR: There may be more information available from
[burl-ct-v440-1:24610] ERROR: the 'qstat -t' command on the Grid Engine tasks.
[burl-ct-v440-1:24610] ERROR: If the problem persists, please restart the
[burl-ct-v440-1:24610] ERROR: Grid Engine PE job
[burl-ct-v440-1:24610] ERROR: The daemon exited unexpectedly with status 1.
We have thought about what happens if a user wants to start two separate jobs and map them to specific nodes with specific rank mappings. Assuming the rank mappings are monotonically increasing from 0 to the maximum rank, how does that map when one has two jobs? It would appear that the two jobs would then get mapped to the exact some locations. One way around this is to set up two sets of ranking within the hostfile and then filter the ones you want for each run. Let me illustrate with an example.
Assume this is the hostfile with my made up rankings and ranking syntax.
# Hostfile names hf
host_a 0,1,4,5
host_b 2,3,6,7
host_c 0,2,4,6
host_d 1,3,5,7
Then, the user could do the following:
orted -universe foobar -persistent -seed
mpirun -universe foobar -np 8 -hostfile hf -host host_a,host_b io_server_job
mpirun -universe foobar -np 8 -host host_c,host_d compute_job
This all seems a little strange. An alternative that was considered was to add the concept of a partition. A partition could be a separate section in a host file or perhaps a different file altogether. It would define a subset of the hostfile. Then, instead of doing all this filtering one could specify the partition name on the mpirun line. ClusterTools had something like this.