Command Parameters of Hyper MPI 25.0.0 and Later Versions
This section describes the command parameters of Hyper MPI 25.0.0 and later versions.
Hyper MPI 25.0.0 and later versions add new command parameters, based on those in Hyper MPI 2.4.3 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_STARS_TLS= |
|
(Optional) Specifies the transmission protocol used by the offloading algorithm. Example: -x UCG_PLANC_STARS_TLS=all |
-x UCG_PLANC_STARS_NET_DEVICES= |
|
(Optional) Specifies the NICs used by the offloading algorithm. Example: -x UCG_PLANC_STARS_NET_DEVICES=all |
-x UCG_PLANC_STARS_SELF_DEVICES= |
|
(Optional) Specifies the loopback device used by the offloading algorithm. Example: -x UCG_PLANC_STARS_SELF_DEVICES=memory |
-x UCG_PLANC_STARS_MAX_RAILS= |
|
(Optional) Specifies the maximum number of links used by the offloading algorithm. Example: -x UCG_PLANC_STARS_MAX_RAILS=1 |
-x UCG_PLANC_STARS_BCAST_ATTR= |
|
(Optional) Specifies a Bcast algorithm parameter. Example: -x UCG_PLANC_STARS_BCAST_ATTR=I:1 |
-x UCG_PLANC_STARS_IBCAST_ATTR= |
|
(Optional) Specifies an IBcast algorithm parameter. Example: -x UCG_PLANC_STARS_IBCAST_ATTR=I:1 |
-x UCG_PLANC_STARS_SCATTERV_ATTR= |
|
(Optional) Specifies a Scatterv algorithm parameter. Example: -x UCG_PLANC_STARS_SCATTERV_ATTR=I:1 |
-x UCG_PLANC_STARS_ISCTTERV_ATTR= |
|
(Optional) Specifies an IScatterv algorithm parameter. Example: -x UCG_PLANC_STARS_ISCATTERV_ATTR=I:1 |
-x UCG_PLANC_STARS_ALLGATHERV_ATTR= |
|
(Optional) Specifies an Allgatherv algorithm parameter. Example: -x UCG_PLANC_STARS_ALLGATHERV_ATTR=I:1 |
-x UCG_PLANC_STARS_IALLGATHERV_ATTR= |
|
(Optional) Specifies an IAllgatherv algorithm parameter. Example: -x UCG_PLANC_STARS_IALLGATHERV_ATTR=I:1 |
-x UCG_PLANC_STARS_ALLTOALLV_ATTR= |
|
(Optional) Specifies an Alltoallv algorithm parameter. Example: -x UCG_PLANC_STARS_ALLTOALLV_ATTR=I:1 |
-x UCG_PLANC_STARS_IALLTOALLV_ATTR= |
|
(Optional) Specifies an IAlltoallv algorithm parameter. Example: -x UCG_PLANC_STARS_IALLTOALLV_ATTR=I:1 |
-x UCG_PLANC_STARS_BARRIER_ATTR= |
|
(Optional) Specifies a Barrier algorithm parameter. Example: -x UCG_PLANC_STARS_BARRIER_ATTR=I:1 |
-x UCG_PLANC_STARS_IBARRIER_ATTR= |
|
(Optional) Specifies an IBarrier algorithm parameter. Example: -x UCG_PLANC_STARS_IBARRIER_ATTR=I:1 |
-x UCG_PLANC_STARS_SCATTERV_KNTREE_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) Specifies a Scatterv algorithm parameter. Example: -x UCG_PLANC_STARS_SCATTERV_KNTREE_DEGREE=2 |
-x UCG_PLANC_STARS_BARRIER_FANIN_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) Specifies a Barrier_faninfanout algorithm parameter. Example: -x UCG_PLANC_STARS_BARRIER_FANIN_DEGREE=4 |
-x UCG_PLANC_STARS_BARRIER_FANOUT_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) Specifies a Barrier_faninfanout algorithm parameter. Example: -x UCG_PLANC_STARS_BARRIER_FANOUT_DEGREE=2 |
-x UCG_PLANC_STARS_SCATTERV_RUN_HPL= |
|
(Optional) Specifies the Scatterv running mode. Example: -x UCG_PLANC_STARS_SCATTERV_RUN_HPL=y |
-x UCG_PLANC_STARS_RC_SEG_SIZE= |
|
(Optional) Specifies the size of the elastic buffer. Example: -x UCG_PLANC_STARS_RC_SEG_SIZE=8256 |
-x UCG_PLANC_STARS_RC_TX_QUEUE_LEN= |
|
(Optional) Specifies the length of the send queue in the QP. Example: -x UCG_PLANC_STARS_RC_TX_QUEUE_LEN=1024 |
-x UCG_PLANC_STARS_RC_TX_MIN_INLINE= |
|
(Optional) Specifies the minimum number of bytes reserved for inline data when the WQE is sent. Example: -x UCG_PLANC_STARS_RC_TX_MIN_INLINE=64 |
-x UCG_PLANC_STARS_RC_TX_INLINE_RESP= |
|
(Optional) Specifies the number of bytes reserved for inline responses when the WQE is sent. Responses that are small enough, such as atomic operations and small reads, are received inline. Example: -x UCG_PLANC_STARS_RC_TX_INLINE_RESP=64 |
-x UCG_PLANC_STARS_RC_HOP_LIMIT= |
|
(Optional) Specifies the IB hop limit/RoCEv2 TTL. The value ranges from 0 to 255. Example: -x UCG_PLANC_STARS_RC_HOP_LIMIT=255 |
-x UCG_PLANC_STARS_RC_PATH_MTU= |
|
(Optional) Specifies the maximum transmission unit (MTU) of the path. default indicates that the optimal MTU is selected for the device. Example: -x UCG_PLANC_STARS_RC_PATH_MTU=1024 |
-x UCG_PLANC_STARS_RC_TIMEOUT= |
|
(Optional) Specifies the transmission timeout interval. Example: -x UCG_PLANC_STARS_RC_TIMEOUT=1ms |
-x UCG_PLANC_STARS_RC_RETRY_COUNT= |
|
(Optional) Specifies the number of transmission retries. Example: -x UCG_PLANC_STARS_RC_RETRY_COUNT=7 |
-x UCG_PLANC_STARS_RC_RNR_TIMEOUT= |
|
(Optional) Specifies the RNR timeout interval. Example: -x UCG_PLANC_STARS_RC_RNR_TIMEOUT=1ms |
-x UCG_PLANC_STARS_RC_RNR_RETRY_COUNT= |
|
(Optional) Specifies the number of RNR retries. Example: -x UCG_PLANC_STARS_RC_RNR_RETRY_COUNT=7 |
-x UCG_PLANC_STARS_RC_TX_CQ_MODERATION= |
|
(Optional) Specifies the maximum number of WQEs that can be sent without requesting completion. Example: -x UCG_PLANC_STARS_RC_TX_CQ_MODERATION=64 |
-x UCG_PLANC_STARS_RC_TX_CQ_LEN= |
|
(Optional) Specifies the length of the TX completion queue. This limits the total number of incomplete signals. Example: -x UCG_PLANC_STARS_RC_TX_CQ_LEN=4096 |
-x UCG_PLANC_STARS_RC_TX_MAX_WR= |
|
(Optional) Limits the number of issued but uncompleted work requests. The actual limit is the smaller value between this value and the TX queue length. –1 indicates no limit. Example: -x UCG_PLANC_STARS_RC_TX_MAX_WR=-1 |
-x UCG_PLANC_STARS_IB_RCACHE_MEM_PRIO= |
|
(Optional) Specifies the priority of the registered cache memory events. Example: -x UCG_PLANC_STARS_IB_RCACHE_MEM_PRIO=1000 |
-x UCG_PLANC_STARS_IB_RCACHE_OVERHEAD= |
|
(Optional) Specifies the overhead of looking up the registered cache. Example: -x UCG_PLANC_STARS_IB_RCACHE_OVERHEAD=180ns |
-x UCG_PLANC_STARS_IB_RCACHE_ADDR_ALIGN= |
|
(Optional) Specifies the registered address alignment cache. The value must be a power of 2 and between 64 and the system page size. Example: -x UCG_PLANC_STARS_IB_RCACHE_ADDR_ALIGN=64 |
-x UCG_PLANC_STARS_IB_REG_METHODS= |
|
(Optional) Specifies the methods for registering MRs with different priorities. Example: -x UCG_PLANC_STARS_IB_REG_METHODS=rcache,odp,direct |
-x UCG_PLANC_STARS_IB_MEM_REG_OVERHEAD= |
|
(Optional) Specifies the memory registration overhead. Example: -x UCG_PLANC_STARS_IB_MEM_REG_OVERHEAD=16us |
-x UCG_PLANC_STARS_IB_MEM_REG_GROWTH= |
|
(Optional) Specifies the memory registration growth rate. Example: -x UCG_PLANC_STARS_IB_MEM_REG_GROWTH=0.1ns |
-x UCG_PLANC_STARS_IB_ETH_PAUSE_ON= |
|
(Optional) Specifies whether to enable pause frames on the Ethernet. A pause frame is a mechanism that temporarily stops data transmission when the Ethernet is congested. If this parameter is set to n, the IB transmission is disqualified because the lossy structure may not perform well when used with RoCE. Example: -x UCG_PLANC_STARS_IB_ETH_PAUSE_ON=y |
-x UCG_PLANC_STARS_IB_ODP_NUMA_POLICY= |
|
(Optional) Sets the NUMA policy for rewriting the ODP zone to avoid additional page migrations.
Example: -x UCG_PLANC_STARS_IB_ODP_NUMA_POLICY=bind |
-x UCG_PLANC_STARS_IB_ODP_MAX_SIZE= |
|
(Optional) Specifies the maximum memory area size for on-demand paging. Example: -x UCG_PLANC_STARS_IB_ODP_MAX_SIZE=16kb |
-x UCG_PLANC_STARS_RC_ROCE_LOCAL_SUBNET= |
|
(Optional) Uses the local IP address and subnet mask of each network device to route RoCEv2 packets. This parameter is used to filter the IB subnet prefix for the port. If this parameter is left empty, no filtering is performed. Example: -x UCG_PLANC_STARS_RC_ROCE_LOCAL_SUBNET=y |
-x UCG_PLANC_STARS_RC_ROCE_LOCAL_SUBNET_PREFIX_LEN= |
|
(Optional) Specifies the length of the subnet prefix used for reachability checks when UCX_IB_ROCE_LOCAL_SUBNET is enabled, in bits. Example: -x UCG_PLANC_STARS_RC_ROCE_LOCAL_SUBNET_PREFIX_LEN=8 |
-x UCG_PLANC_STARS_SDMA_BW= |
|
(Optional) Specifies the SDMA bandwidth, which is used to control the packet size of the SDMA link during data transmission. This parameter takes effect only in the multi-rail scenario. Example: -x UCG_PLANC_STARS_SDMA_BW=10000MBs |
-x UCX_RC_VERBS_TX_MIN_SGE= |
|
(Optional) Specifies the number of SG entries retained in the WQE when sending over RC. Example: -x UCX_RC_VERBS_TX_MIN_SGE=2 |
-x UCX_UD_VERBS_TX_MIN_SGE= |
|
(Optional) Specifies the number of SG entries retained in the WQE when sending over UD. Example: -x UCX_UD_VERBS_TX_MIN_SGE=1 |
-x UCX_SDMA_SHARED_CHANNEL_MODE= |
|
(Optional) After this function is enabled, the SDMA can use the shared channel. Example: -x UCX_SDMA_SHARED_CHANNEL_MODE=y |
-x UCG_PLANC_STARS_STREAM_DEPTH= |
|
(Optional) Specifies the depth of the offloading queue. Example: -x UCG_PLANC_STARS_STREAM_DEPTH=4096 |
-x UCG_PLANC_UCX_ALLGATHERV_DEFAULT_POLICY= |
|
(Optional) Specifies the optimization algorithm used by the Allgatherv operator. (This parameter may not be applicable in some environments.) Example: -x UCG_PLANC_UCX_ALLGATHERV_DEFAULT_POLICY=y |
-x UCG_PLANC_UCX_ALLTOALLV_DEFAULT_POLICY= |
|
(Optional) Specifies the optimization algorithm used by the Alltoallv operator. (This parameter may not be applicable in some environments.) Example: -x UCG_PLANC_UCX_ALLTOALLV_DEFAULT_POLICY=y |