Yari,
Thanks for the pointer. I looked at the the tuning guide, it is good, but targeted for ethernet over IB performance and doesn't delve much into details of IB driver settings, although there is a section about autotuning.
But you may have some more, see the list of params here for kernel modules:
We see that tuning/increasing the default params in /sys/module/mlx4_core/parameters resolve issues, we even see some performance boost for our use case. (when we increase log_num_mtt, log_mtts_per_seg).
We really need to understand how to obtain max performance tuning these and other driver params. I doubt the auto-tuners would resolve it for us..
ll /sys/module/mlx4_core/parameters/
-rw-r--r-- 1 root root 4096 Mar 11 17:23 block_loopback
-rw-r--r-- 1 root root 4096 Mar 11 17:23 debug_level
-rw-r--r-- 1 root root 4096 Mar 11 17:23 enable_pre_t11_mode
-r--r--r-- 1 root root 4096 Mar 11 17:23 enable_qinq
-r--r--r-- 1 root root 4096 Mar 11 17:23 enable_qos
-r--r--r-- 1 root root 4096 Mar 11 17:23 fast_drop
-r--r--r-- 1 root root 4096 Mar 11 17:23 high_rate_steer
-rw-r--r-- 1 root root 4096 Mar 11 17:23 internal_err_reset
-r--r--r-- 1 root root 4096 Mar 11 17:23 log_mtts_per_seg
-r--r--r-- 1 root root 4096 Mar 11 17:23 log_num_cq
-r--r--r-- 1 root root 4096 Mar 11 17:23 log_num_mac
-r--r--r-- 1 root root 4096 Mar 11 17:23 log_num_mcg
-r--r--r-- 1 root root 4096 Mar 11 17:23 log_num_mgm_entry_size
-r--r--r-- 1 root root 4096 Mar 11 17:23 log_num_mpt
-r--r--r-- 1 root root 4096 Mar 11 17:23 log_num_mtt
-r--r--r-- 1 root root 4096 Mar 11 17:23 log_num_qp
-r--r--r-- 1 root root 4096 Mar 11 17:23 log_num_srq
-r--r--r-- 1 root root 4096 Mar 11 17:23 log_rdmarc_per_qp
-r--r--r-- 1 root root 4096 Mar 11 17:23 msi_x
-r--r--r-- 1 root root 4096 Mar 11 17:23 pfcrx
-r--r--r-- 1 root root 4096 Mar 11 17:23 pfctx
-r--r--r-- 1 root root 4096 Mar 11 17:23 probe_vf
-r--r--r-- 1 root root 4096 Mar 11 17:23 set_4k_mtu
-r--r--r-- 1 root root 4096 Mar 11 17:23 sr_iov
-r--r--r-- 1 root root 4096 Mar 11 17:23 use_prio
ll /sys/module/mlx4_en/parameters
total 0
-r--r--r-- 1 root root 4096 Mar 11 17:40 enable_sys_tune
-r--r--r-- 1 root root 4096 Mar 11 17:40 inline_thold
-r--r--r-- 1 root root 4096 Mar 11 17:40 mem_node
-r--r--r-- 1 root root 4096 Mar 11 17:40 num_rx_rings
-r--r--r-- 1 root root 4096 Mar 11 17:40 udp_rss
Thanks,
Mous.