IntelĀ® MPI Library Reference Manual for Linux* OS
Set the number of connection adapters.
I_MPI_OFA_NUM_ADAPTERS=<arg>
<arg> |
Define the maximum number of connection adapters used |
>0 |
Use the specified number of adapters. The default value is 1 |
Set the number of the adapters that are used. If the number is greater than the available number of adapters, all the available adaptors are used.
Set the name of adapter that is used.
I_MPI_OFA_ADAPTER_NAME=<arg>
<arg> |
Define the name of adapter |
Name |
Use the specified adapter. By default, any adapter can be used |
Set the name of adaptor to be used. If the adapter with specified name does not exist, the library returns an error. This has effect only if I_MPI_OFA_NUM_ADAPTERS=1.
Set the number of used ports on each adapter.
I_MPI_OFA_NUM_PORTS=<arg>
<arg> |
Define the number of ports that are used on each adapter |
> 0 |
Use the specified number of ports. The default value is 1 |
Set the number of used ports on each adaptor. If the number is greater than the available number of ports, all the available ports are used.
I_MPI_OFA_NUM_RDMA_CONNECTIONS
Set the maximum number of connections that can use the rdma exchange protocol.
I_MPI_OFA_NUM_RDMA_CONNECTIONS=<num_conn>
<num_conn> |
Define the maximum number of connections that can use the rdma exchange protocol |
>= 0 |
Create the specified number of connections that use the rdma exchange protocol. All other processes use the send/ receive exchange protocol |
-1 |
Create log2(number of processes) rdma connections |
>= number of processes |
Create rdma connections for all processes. This is the default value |
There are two exchange protocols between two processes: send/receive and rdma. This environment variable specifies the maximum amount of connections that use rdma protocol.
RDMA protocol is faster but requires more resources. For a large application, you can limit the number of connections that use the rdma protocol so that only processes that actively exchange data use the rdma protocol.
Set the number of messages that a process should receive before switching this connection to RDMA exchange protocol.
I_MPI_OFA_SWITCHING_TO_RDMA=<number>
<number> |
Define the number of messages that the process receives before switching to use the rdma protocol |
>= 0 |
If this process receives <number> of messages, start using the rdma protocol |
Count the number of messages received from the specific process. The connection achieved the specified number tries to switch to rdma protocol for exchanging with that process. The connection will not switch to rdma protocol if the maximum number of connections that use the rdma exchange protocol defined in I_MPI_OFA_NUM_RDMA_CONNECTIONS has been reached.
Set the method of choosing rails for short messages.
I_MPI_OFA_RAIL_SCHEDULER=<arg>
<arg> |
Mode selector |
ROUND_ROBIN |
Next time use next rail |
FIRST_RAIL |
Always use the first rail for short messages |
PROCESS_BIND |
Always use the rail specific for process |
Set the method of choosing rails for short messages. The algorithms are selected according to the following scheme:
In the ROUND_ROBIN mode, the first message is sent using the first rail; the next message is sent using the second rail, and so on.
In the FIRST_RAIL mode, the first rail is always used for short messages.
In the PROCESS_BIND mode, the process with the smallest rank uses the first rail, and the next uses the second rail.
Turn on/off the memory registration cache.
I_MPI_OFA_TRANSLATION_CACHE=<arg>
<arg> |
Binary indicator |
enable | yes | on | 1 |
Turn on the memory registration cache. This is the default |
disable | no | off | 0 |
Turn off the memory registration cache |
Set this environment variable to turn on/off the memory registration cache.
The cache substantially increases performance, but may lead to correctness issues in certain situations. See product Release Notes for further details.
I_MPI_OFA_TRANSLATION_CACHE_AVL_TREE
Enable/disable the AVL tree* based implementation of the RDMA translation cache.
I_MPI_OFA_TRANSLATION_CACHE_AVL_TREE=<arg>
<arg> |
Binary indicator |
enable | yes | on | 1 |
Turn on the AVL tree based RDMA translation cache |
disable | no | off | 0 |
Turn off the AVL tree based RDMA translation cache. This is the default value |
Set this environment variable to enable the AVL tree based implementation of RDMA translation cache in the OFA path. When the search in RDMA translation cache handles over 10,000 elements, the AVL tree based RDMA translation cache is faster than the default implementation.
Control the library to create queue pairs (QPs) dynamically.
I_MPI_OFA_DYNAMIC_QPS=<arg>
<arg> |
Binary indicator |
enable | yes | on | 1 |
Create QPs dynamically. This is the default value if the number of processes is larger than or equal 2,000 |
disable | no | off | 0 |
Create all QPs during the initial stage. This is the default value if the number of processes is less than 2,000 |
Set this environment variable to turn on dynamic creation of QPs.
Set the size of the packet used for sending.
I_MPI_OFA_PACKET_SIZE=<arg>
<arg> |
Define the size of packet in bytes |
> 0 |
Use the specified packet size. The default value is 8192 |
Set the packet size in bytes. If the number is negative, the size is set to 8.
Set the name of the used OFA library.
I_MPI_OFA_LIBRARY=<arg>
<arg> |
Define the name of the OFA library |
Name |
Use the specified library. By default, the name is libibverbs.so |
Set the name of the InfiniBand* (IB*) library. If the library with the specified name does not exist, an error is returned.
Define the nonstandard template for port connections.
I_MPI_OFA_NONSWITCH_CONF=<arg>
<arg> |
Define the template for port connections |
Name |
Use the specified template |
The nodes in clusters are normally connected so that porti of a node can access porti of all other nodes. Use this environment variable if ports are not connected in this way. The following example is the template format:
host1@port11#port12#...#host2@port21#port22....
Portij defines the port used to send from hosti to hostj
For example:
node1@1#1#2#node2@2#1#1#node3@1#2#1#
This sample specifies the following configuration:
Port1 of node1 connected to port2 of node2
Port2 of node1 connected to port1 of node3
Port1 of node2 connected to port2 of node3
Port2 of node2 connected to port1 of node2
Port1 of node3 connected to port2 of node1
Port2 of node3 connected to port1 of node2
Port1 is always used to communicate with itself (loopback).