Strix Halo Wiki
Attachments
History
Blame
View Source
A - Z
Changelog
Documentation
About An Otter Wiki
Toggle dark mode
Login
Menu
Home
Buyer's Guide
AI Guides
General Guides
Boards
PCs
GitHub Mirror
Discord Server
Strix Halo Toolboxes
Page Index
AI
AI Capabilities Overview
Clustering
Clustering with RDMA
llamacpp-performance
llamacpp-with-ROCm
vLLM
AI
Clustering with RDMA
355af7
Commit
355af7
2026-05-06 20:44:58
Lorphos
: hide problematic vlan tuning
AI/Clustering_with_RDMA.md
..
@@ 61,11 61,12 @@
[ 1943.886040] mlx4_core 0000:c3:00.0 ibp195s0: Port: 1 Link INIT
[ 1943.941515] mlx4_core 0000:c3:00.0 ibp195s0: Port: 1 Link ACTIVE
```
+
<!-- this is problematic
To enable performance optimized steering (and surrender VLAN support), edit
`/etc/modprobe.d/mlx4.conf` and add this line:
```options mlx4_core log_num_mgm_entry_size=-7
```
-
as mentioned in the [driver documentation](https://doc.dpdk.org/guides/nics/mlx4.html).
+
as mentioned in the [driver documentation](https://doc.dpdk.org/guides/nics/mlx4.html). -->
Install needed packages on both PCs running Fedora 43:
```$ sudo dnf install rdma-core libibverbs-utils mstflint infiniband-diags perftest
0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9