Command Parameters of Hyper MPI 1.3.* and Later Versions
This section describes the command parameters of Hyper MPI 1.3.* and later versions.
When using Hyper MPI command parameters to specify algorithms, you can refer to Algorithms for the meaning of specific parameter values.
Parameter |
Value |
Description |
|---|---|---|
-x UCG_PLANC_UCX_ALLREDUCE_FANOUT_INTER_DEGREE= |
NOTE:
You are advised to set this parameter to a positive integer ranging from 2 to the number of job nodes, for example, 2 to 8.
|
(Optional) This parameter is used in Allreduce algorithms 5, 6, 7, and 8. Example: -x UCG_PLANC_UCX_ALLREDUCE_FANOUT_INTER_DEGREE=8 |
-x UCG_PLANC_UCX_ALLREDUCE_FANIN_INTER_DEGREE= |
NOTE:
You are advised to set this parameter to a positive integer ranging from 2 to the number of job nodes, for example, 2 to 8.
|
(Optional) This parameter is used in Allreduce algorithms 5, 6, 7, and 8. Example: -x UCG_PLANC_UCX_ALLREDUCE_FANIN_INTER_DEGREE=8 |
-x UCG_PLANC_UCX_ALLREDUCE_FANOUT_INTRA_DEGREE= |
NOTE:
You are advised to set this parameter to a positive integer ranging from 2 to the number of processes involved in communication, for example, 2 to 8.
|
(Optional) This parameter is used in Allreduce algorithms 7 and 8. Example: -x UCG_PLANC_UCX_ALLREDUCE_FANOUT_INTRA_DEGREE=2 |
-x UCG_PLANC_UCX_ALLREDUCE_FANIN_INTRA_DEGREE= |
NOTE:
You are advised to set this parameter to a positive integer ranging from 2 to the number of processes involved in communication, for example, 2 to 8.
|
(Optional) This parameter is used in Allreduce algorithms 7 and 8. Example: -x UCG_PLANC_UCX_ALLREDUCE_FANIN_INTRA_DEGREE=4 |
-x UCG_PLANC_UCX_BARRIER_FANOUT_INTER_DEGREE= |
NOTE:
You are advised to set this parameter to a positive integer ranging from 2 to the number of job nodes, for example, 2 to 8.
|
(Optional) This parameter is used in Barrier algorithms 4, 5, 6, and 7. Example: -x UCG_PLANC_UCX_BARRIER_FANOUT_INTER_DEGREE=8 |
-x UCG_PLANC_UCX_BARRIER_FANIN_INTER_DEGREE= |
NOTE:
You are advised to set this parameter to a positive integer ranging from 2 to the number of job nodes, for example, 2 to 8.
|
(Optional) This parameter is used in Barrier algorithms 4, 5, 6, and 7. Example: -x UCG_PLANC_UCX_BARRIER_FANIN_INTER_DEGREE=8 |
-x UCG_PLANC_UCX_BARRIER_FANOUT_INTRA_DEGREE= |
NOTE:
You are advised to set this parameter to a positive integer ranging from 2 to the number of processes involved in communication, for example, 2 to 8.
|
(Optional) This parameter is used in Barrier algorithms 6 and 7. Example: -x UCG_PLANC_UCX_BARRIER_FANOUT_INTRA_DEGREE=2 |
-x UCG_PLANC_UCX_BARRIER_FANIN_INTRA_DEGREE= |
NOTE:
You are advised to set this parameter to a positive integer ranging from 2 to the number of processes involved in communication, for example, 2 to 8.
|
(Optional) This parameter is used in Barrier algorithms 6 and 7. Example: -x UCG_PLANC_UCX_BARRIER_FANIN_INTRA_DEGREE=8 |
-x UCG_PLANC_UCX_BCAST_ATTR= |
|
(Optional) Specifies a Bcast algorithm parameter. Example: -x UCG_PLANC_UCX_BCAST_ATTR=I:4 |
-x UCG_PLANC_UCX_ALLREDUCE_ATTR= |
|
(Optional) Specifies an Allreduce algorithm parameter. Example: -x UCG_PLANC_UCX_ALLREDUCE_ATTR=I:7 |
-x UCG_PLANC_UCX_BARRIER_ATTR= |
|
(Optional) Specifies a Barrier algorithm parameter. Example: -x UCG_PLANC_UCX_BARRIER_ATTR=I:1 |
-x UCG_PLANC_UCX_ALLTOALLV_ATTR= |
|
(Optional) Specifies an Alltoallv algorithm parameter. Example: -x UCG_PLANC_UCX_ ALLTOALLV_ATTR=I:1 |
-x UCG_PLANC_UCX_ALLGATHERV_ATTR= |
|
(Optional) Specifies an Allgatherv algorithm parameter. Example: -x UCG_PLANC_UCX_ ALLGATHERV_ATTR=I:3 |
-x UCG_PLANC_UCX_SCATTERV_ATTR= |
|
(Optional) Specifies a Scatterv algorithm parameter. Example: -x UCG_PLANC_UCX_ SCATTERV_ATTR=I:1 |
-x UCG_PLANC_UCX_GATHERV_ATTR= |
|
(Optional) Specifies a Gatherv algorithm parameter. Example: -x UCG_PLANC_UCX_ GATHERV_ATTR=I:1 |
-x UCG_PLANC_UCX_BCAST_NA_KNTREE_INTER_DEGREE= |
NOTE:
You are advised to set this parameter to a positive integer ranging from 2 to the number of job nodes, for example, 2 to 8.
|
(Optional) This parameter is used in Bcast algorithms 3 and 4. Example: -x UCG_PLANC_UCX_BCAST_NA_KNTREE_INTER_DEGREE=8 |
-x UCG_PLANC_UCX_BCAST_NA_KNTREE_INTRA_DEGREE= |
NOTE:
You are advised to set this parameter to a positive integer ranging from 2 to the number of processes involved in communication, for example, 2 to 8.
|
(Optional) This parameter is used in Bcast algorithm 4. Example: -x UCG_PLANC_UCX_BCAST_NA_KNTREE_INTRA_DEGREE=2 |
-x UCG_PLANC_UCX_REDUCE_KNTREE_DEGREE= |
NOTE:
You are advised to set this parameter to a positive integer ranging from 2 to the number of processes involved in communication, for example, 2 to 8.
|
(Optional) This parameter is used in Allreduce algorithms 2, 3, 5, 6, 7, and 8. Example: -x UCG_PLANC_UCX_REDUCE_KNTREE_DEGREE=2 |
-x UCG_PLANC_UCX_SCATTERV_KNTREE_DEGREE= |
NOTE:
You are advised to set this parameter to a positive integer ranging from 2 to the number of processes involved in communication, for example, 2 to 8.
|
(Optional) This parameter is used in Scatterv algorithm 2. Example: -x UCG_PLANC_UCX_SCATTERV_KNTREE_DEGREE=2 |
-x UCG_PLANC_UCX_SCATTERV_MIN_SEND_BATCH= |
|
(Optional) Specifies the minimum boundary value of the TX batch processing mode for Scatterv algorithm 1. Example: -x UCG_PLANC_UCX_SCATTERV_MIN_SEND_BATCH=1 |
-x UCG_PLANC_UCX_SCATTERV_MAX_SEND_BATCH= |
|
(Optional) Specifies the maximum boundary value of the TX batch processing mode for Scatterv algorithm 1. Example: -x UCG_PLANC_UCX_SCATTERV_MIN_SEND_BATCH=1 |
-x UCG_PLANC_UCX_NPOLLS= |
|
(Optional) Specifies the number of UCP progress polling cycles for the P2P request test. Example: -x UCG_PLANC_UCX_NPOLLS=3 NOTE:
In a unidirectional request involving atomic operations (for example, osu_get_acc_latency/osu_acc_latency in the OSU benchmark), you can reduce the value (for example, set it to 0) to improve performance. |
-x UCG_PLANC_UCX_USE_OOB= |
|
(Optional) Reuses the OMPI PML UCX link. Example: -x UCG_PLANC_UCX_USE_OOB=yes |
-x UCG_USE_MT_MUTEX= |
|
(Optional) Specifies the implementation mode of the multi-thread lock in the UCG. Example: -x UCG_USE_MT_MUTEX=n |
-x UCG_LOG_LEVEL= |
NOTE:
Log levels are ordered as: fatal > error > warn > info > debug > trace. The lower the level, the more detailed the log output. For example, if the level is set to warn, the system will output logs at the fatal, error, and warn levels.
|
(Optional) Specifies the UCG log level. Example: -x UCG_LOG_LEVEL=warn |
-x UCG_PLANC_UCX_PLANM= |
|
(Optional) Specifies whether to load the closed-source libraries. Example: -x UCG_PLANC_UCX_PLANM=all |
-x UCG_PLANC_UCX_ALLREDUCE_NUM_NAP_GROUP= |
|
(Optional) This parameter is used in Allreduce algorithm 11. Example: -x UCG_PLANC_UCX_ALLREDUCE_NUM_NAP_GROUP=1 |
-x UCG_PLANC_UCX_BARRIER_NUM_NAP_GROUP= |
|
(Optional) This parameter is used in Barrier algorithm 10. Example: -x UCG_PLANC_UCX_BARRIER_NUM_NAP_GROUP=1 |
-x UCG_PLANC_UCX_BCAST_ESBT_BLOCKS= |
|
(Optional) This parameter is used in Bcast algorithm 9. Example: -x UCG_PLANC_UCX_BCAST_ESBT_BLOCKS=0 |
-x UCG_PLANC_UCX_ALLTOALLV_MIN_SEND_BATCH= |
|
(Optional) Specifies the minimum boundary value of the TX batch processing mode for Alltoallv algorithm 1. Example: -x UCG_PLANC_UCX_ALLTOALLV_MIN_SEND_BATCH=0 |
-x UCG_PLANC_UCX_ALLTOALLV_MAX_SEND_BATCH= |
|
(Optional) Specifies the maximum boundary value of the TX batch processing mode for Alltoallv algorithm 1. Example: -x UCG_PLANC_UCX_ALLTOALLV_MAX_SEND_BATCH=1 |
-x UCG_PLANC_UCX_REDUCE_CONSISTENCY= |
|
(Optional) Specifies whether to ensure the consistency of Allreduce results in UCG. Example: -x UCG_PLANC_UCX_REDUCE_CONSISTENCY=n |
--mca coll_ucg_priority |
|
(Optional) Specifies the priority for Hyper MPI to invoke the UCG module. A larger value indicates a higher priority. Example: --mca coll_ucg_priority 90 |
--mca coll_ucg_verbose |
|
(Optional) Specifies the log verbosity of the UCG component. Example: --mca coll_ucg_verbose 2 |
--mca coll_ucg_max_rcache_size |
|
(Optional) Specifies the request cache size of the COLL UCG component. Example: --mca coll_ucg_max_rcache_size 10 |
--mca coll_ucg_disable_coll |
|
(Optional) Specifies the collective operations that are disabled after Hyper MPI optimization when a Hyper MPI job is running. Example: --mca coll_ucg_disable_coll barrier,bcast |
--mca topo_tpm_enable |
|
(Optional) Specifies whether to enable the process layout function. Example: --mca topo_tpm_enable 0 |
--mca pmix_pmix3x_wait_time |
|
(Optional) Specifies the duration for the client process to wait for the processing result from the server. Example: --mca pmix_pmix3x_wait_time 5 |
-x UCX_UD_VERBS_MIN_PENDING_TICK |
|
(Optional) Specifies the minimum scheduler suspension time to prevent false timeout reporting due to scheduler suspension under the UD VERBS protocol. Example: -x UCX_UD_VERBS_MIN_PENDING_TICK=2s |
-x UCX_UD_MLX5_MIN_PENDING_TICK |
|
(Optional) Specifies the minimum scheduler suspension time to prevent false timeout reporting due to scheduler suspension under the UD MLX5 protocol. Example: -x UCX_UD_MLX5_MIN_PENDING_TICK=2s |
-x UCX_MIN_PENDING_TIME |
|
(Optional) Specifies the minimum scheduler suspension time to prevent false timeout reporting due to scheduler suspension during UCP request timeout detection. Example: -x UCX_MIN_PENDING_TIME=2s |
-x UCX_REQ_TIMEOUT |
|
(Optional) Specifies the timeout threshold for UCP request timeout detection. When the duration of a point-to-point communication request that is not completed exceeds the threshold, a timeout alarm is generated. Example: -x UCX_REQ_TIMEOUT=100s |