Known Issues

NVIDIA Scalable Hierarchical Aggregation and Reduction Protocol (SHARP) Rev 3.6.0

Internal Reference Number

Issues

3478803

Description: Getting topology info (sharp_cmd topology), fails when executed from the mgmt host.

Workaround: It is possible to run it from different hosts, or add the following environment variable: SHARP_ALLOW_SM_PORT=1

Keywords: SHARP topology API

Discovered in Version: 3.5.0

3340353

Description: When reconfiguring a standby management host to operate as a compute host, it will not be able to run SHARP jobs unless sharp_am is restarted.

In case that a host runs the SM process, it will automatically be detected by the master SM as a standby SM and be reported as a standby management host.

Note that restart is not required if ignore_sm_guids is set to FALSE.

Workaround: N/A

Keywords: Slave; compute host; ignore_sm_guids

Discovered in Version: 3.3.0

3371820

Description: Congestion Control cannot be configured on the same SLs used by sharp_am.

Workaround: N/A

Keywords: Congestion control; SL

Discovered in Version: 3.3.0

3438393

Description: When operating in the following configuration mode, resource limitation is ignored and no limit is set to any application:

Dynamic trees allocation is used; Quasi Fat Tree (QFT)-oriented logic is used; and reservation_mode is on.

Workaround: N/A

Keywords: Dynamic trees allocation; QFT; resource limitation

Discovered in Version: 3.3.0

3305335

Description: When running mpirun with multiple groups, the following error message might be received:

[error] - AM QPAlloc confirm QP MAD response status 0x1c00

This message is received due to to the fact that multiple unserialized MAD requests are run in parallel.

Workaround: Set the SHARP_COLL_SERIALIZE_MADS environment variable to TRUE when running mpirun.

Keywords: mpirun; SHARP_COLL_SERIALIZE_MADS

Discovered in Version: 3.2.0

3225401

Description: Dynamic trees creation feature does not support a case in which all root switches are down and restarted. If such a scenario takes place, sharp_am should be restarted once the root switches are up and running.

Workaround: N/A

Keywords: Aggregation Manager; sharp_am; dynamic trees

Discovered in Version: 3.1.0

3237831

Description: SHARP does not support reassignment of LID values.

In case LID reassignment is desired, make sure to stop all SHARP jobs, reassign LIDs via OpenSM, and restart sharp_am once the reassignment is done.

Workaround: N/A

Keywords: Aggregation Manager; OpenSM

Discovered in Version: 3.1.0

3048427

Description: In the case that a switch split mode is modified (off/on), sharp_am does not handle the new number of supported ports unless it is restarted.

Workaround: Restart sharp_am after changing a switch split mode definition.

Keywords: Aggregation Manager; split mode

Discovered in Release: 2.7.0

3051699

Description: Changing the configuration of SHARP switch ports using device_configuration_file does not take effect on disconnected split ports. If these ports are connected later, they will remain with their default configuration.

Workaround: If the new configuration is desired for the split ports, make sure to restart the Aggregation Manager after connecting a split port to a host.

Keywords: Aggregation Manager; split port

Discovered in Release: 2.7.0

3051924

Description: Adding or replacing non-leaf switches is currently not supported by Aggregation Manager for Dragonfly+ topologies.

Workaround: Restart Aggregation Manager after the Subnet Manager completes fabric reconfiguration followed by the fabric changes.

Keywords: Fabric extension; Aggregation Manager; AM

Discovered in Release: 2.7.0

-

Description: On multi PKEY environment, UCX in SHARP can use only the default PKEY (PKEY at index 0).

Workaround: Use sockets for communication over non-default PKEY.

Keywords: Configuration, SMX, UCX, PKEY

Discovered in Release: 2.4.3

1307124

Description: Begin Job requests with virtual ports might be rejected until fabric virtualization info file is parsed.

Workaround: Wait for AM to discover virtual ports before sending Begin Job requests.

Keywords: Aggregation Manager, Socket Direct, Virtual Ports

Discovered in Release: 1.5.3

1193629

Description: Configuring sharp_am as daemon is not possible when installing from RPM into non-default location.

Workaround: Configure daemon manually.

Keywords: Configuration

Discovered in Release: 1.5.3

1307108

Description: Discovering a new Aggregation Node (AN) found on the shortest path between two ANs might invalidate the existing path.

Workaround: Restart Aggregation Manager after the Subnet Manager completes fabric reconfiguration followed by the fabric changes.

Keywords: Aggregation Manager, Aggregation Node

Discovered in Release: 1.5.3

-

Description: Aggregation Manager High Availability is currently not supported in HPCX/MLNX OFED packages. Therefore, only a single instance of Aggregation Manager can run in the IB fabric.

Workaround: Use Aggregation Manager in UFM.

Keywords: Aggregation Manager

-

Description: Aggregation manager should run on the same Host where the Master Subnet Manager (SM) is running.

Workaround: N/A

Keywords: Aggregation Manager

-

Description: In case of HPCX/MLNX OFED packages, upon Subnet Manager handover/failover, another instance of Aggregation Manager should be started on the Host where the new Master SM is running

Workaround: Use Aggregation Manager in UFM.

Keywords: Aggregation Manager

-

Description: Aggregation Manager should be started after completion of fabric configuration by the Subnet Manager.

Workaround: N/A

Keywords: Aggregation Manager

-

Description: Only Fat-Tree, Quasi-Fat-Tree, Hypercube and Dragonfly+ topologies are supported by the Aggregation Manager.

Workaround: N/A

Keywords: Fabric Topology

-

Description: Only IB fabrics where all compute nodes are connected to NVIDIA SHARP capable switches are supported by the Aggregation Manager.

Workaround: Manually configure mapping between the compute port and the Aggregation Node.

Keywords: Fabric Topology

-

Description: Upon changes in configuration file beyond parameters in 3.3, Aggregation Manager should be restarted to deploy new configuration.

Workaround: N/A

Keywords: Configuration

3686321

Description: When upgrading UFM from previous versions to UFM 6.15.x, sharp_am persistent directory as mentioned in the configuration file directs to a path that does not exist.

This leads to failure in saving reservation and job information, so in case of a restart of sharp_am, it won’t be able to retrieve required information and return to its previous state.

Workaround: Edit the file /opt/ufm/files/conf/sharp/sharp_am.cfg, modify the parameter persistent_dir, to direct to the path: /opt/ufm/files/conf/sharp/jobs.

Make sure this part does exist.

Keywords: sharp_am, UFM, upgrade

Discovered in Version: 3.5.0

© Copyright 2023, NVIDIA. Last updated on Feb 8, 2024.