I was trying to test the msg rate.
With netperf, use 64 points to test and i got this. It’s much far away from the MAX.
What can I do now?
Hello 1054919371,
Thank you for posting your inquiry to the NVIDIA Developer Forums.
Some troubleshooting/tuning recommendations follow - if these do not resolve your issue, and you have valid support entitlement, we highly recommend engaging NVEX Support for further assistance on this issue.
1. System Tuning
-
Use tuned-adm profile network-throughput to set optimal network parameters
-
Disable IRQ balancing: systemctl stop irqbalance.service
-
Optimize CPU governor for performance
2. Network Interface Optimization
-
Increase interface queue size:
ethtool -G <interface> rx 8192 tx 8192
-
Disable adaptive coalescing:
ethtool -C <interface> adaptive-rx off adaptive-tx off
-
Optimize interrupt moderation:
ethtool -C <interface> rx-usecs 0 rx-frames 0
-
Set appropriate queue count:
ethtool -L <interface> combined 31
3. IRQ Affinity Optimization
- Use set_irq_affinity_cpulist.sh to assign IRQs to specific CPUs
4. RSS Configuration for BlueField-3
-
Implement proper RSS rules to distribute traffic across multiple queues
-
Ensure 1:1 ratio between streams and RX queues
-
Use multiple IP addresses for better queue utilization
-
Apply TC rules with set_tc_rules.sh p0 <interface_name>
5. Netperf Testing Methodology
-
Start with a single stream baseline test, then scale up
-
Use:
netperf -H <server_ip> -t UDP_STREAM -l 60 -- -m 64
-
For multiple streams, launch parallel netperf instances
-
Test with different message sizes (8, 64, 128, 512 bytes)
Best regards,
NVIDIA Enterprise Experience