Site icon New Generation Enterprise Linux

Stop the kernel from dropping packets when your NIC buffers overflow

Performance Monitoring & Tuning

Stop the kernel from dropping packets when your NIC buffers overflow

Technical Briefing | 11/17/2026

You spend all morning tuning your application stack, tweaking thread pools and garbage collection, but the throughput still flatlines. It is not the app code. It is not even the disk I/O. Your network interface card is drowning, and the kernel is silently tossing packets into the void before they even reach your stack. I have seen this drive teams mad because the server looks idle while the request latency spikes.

Why ethtool -S is the only tool that matters here

Most people check ifconfig or ip -s and call it a day. But those tools only show you what the kernel counted on the stack side. To see if your physical NIC is actually overwhelmed, you need the hardware-specific counters. Most drivers report dropped packets right in the ring buffer statistics. If you see rx_missed_errors or similar keys climbing, your NIC is literally telling you it is too busy to catch what is coming off the wire.

ethtool -S eth0 | grep -E 'drop|missed|discard'
  • Check rx_no_dma_resources if you are dealing with high-bandwidth traffic
  • Look for rx_missed_errors to identify physical buffer saturation on the card
  • Monitor tx_dropped to see if the interface queue is too small for your burst patterns

Fixing the ring buffer bottleneck

If you confirm the drops, you have two choices. You can increase the ring buffer size via ethtool -G, which helps with bursty traffic by providing a larger cushion. But be careful. If you set this too high, you might increase latency because the kernel will spend more time processing a bloated buffer. If you cannot solve it with tuning, you are hitting the physical limits of the NIC and need to look into RSS or queue distribution.

Next time your app starts feeling sluggish during a traffic spike, check the hardware counters before you waste hours profiling code that is perfectly healthy. It is usually the pipe, not the water.

Linux Admin Automation  |  © www.ngelinux.com  |  11/17/2026
0 0 votes
Article Rating
Exit mobile version