SysTutorials

Tuning Linux UDP Buffer Sizes to Prevent Packet Loss

· Eric Ma

UDP packet loss on Linux commonly stems from undersized kernel buffers. When packets arrive faster than the application can process them, the kernel drops excess datagrams silently. Unlike TCP, UDP offers no retransmission, so these losses are permanent. Proper buffer tuning is essential for DNS servers, NTP, monitoring systems, video streaming, and any high-throughput UDP service.

Why UDP Packets Get Lost

The kernel maintains separate receive and transmit buffers for each socket. When a UDP packet arrives and the receive buffer is full, the kernel discards it. This happens even when the application is running normally—the issue is purely a buffering mismatch between network arrival rate and application consumption rate.

Checking Current Buffer Configuration

View your system’s current socket buffer limits:

sysctl net.core.rmem_max net.core.rmem_default
sysctl net.core.wmem_max net.core.wmem_default

Default values on most modern distributions are modest (typically 128 KB to 256 KB). Check what a specific process is using with /proc:

cat /proc/net/udp

The RxQueue column shows how full the receive buffer currently is. If you regularly see non-zero values, packets are being dropped.

To monitor in real-time with ss:

ss -tnmu | head -20

Increasing Buffer Sizes Temporarily

For immediate testing (persists until reboot):

sudo sysctl -w net.core.rmem_max=26214400
sudo sysctl -w net.core.rmem_default=26214400
sudo sysctl -w net.core.wmem_max=26214400
sudo sysctl -w net.core.wmem_default=26214400

This sets both receive and transmit buffers to 25 MB. Changes take effect immediately.

Permanent Configuration

Create or edit /etc/sysctl.d/99-udp-buffers.conf:

sudo tee /etc/sysctl.d/99-udp-buffers.conf > /dev/null <<EOF
net.core.rmem_max=26214400
net.core.rmem_default=26214400
net.core.wmem_max=26214400
net.core.wmem_default=26214400
EOF

Apply the changes:

sudo sysctl -p /etc/sysctl.d/99-udp-buffers.conf

Verify the new values:

sysctl net.core.rmem_max
cat /proc/sys/net/core/rmem_max

Application-Level Buffer Configuration

System limits set the maximum, but applications must explicitly request larger buffers. Most UDP applications have configuration options or allow socket buffer settings via SO_RCVBUF and SO_SNDBUF socket options.

In C:

int rcvbuf = 26214400;
setsockopt(sock, SOL_SOCKET, SO_RCVBUF, &rcvbuf, sizeof(rcvbuf));

In Python:

import socket
sock = socket.socket(socket.AF_INET, socket.SOCK_DGRAM)
sock.setsockopt(socket.SOL_SOCKET, socket.SO_RCVBUF, 26214400)

The kernel enforces the system rmem_max as a hard cap—requesting a larger buffer than the system limit will be silently capped.

In Go:

import "net"
conn, _ := net.ListenUDP("udp", &net.UDPAddr{Port: 5353})
file, _ := conn.File()
defer file.Close()
syscall.SetsockoptInt(int(file.Fd()), syscall.SOL_SOCKET, syscall.SO_RCVBUF, 26214400)

Right-Sizing Buffer Values

Buffer size recommendations depend on your workload:

Choose conservatively based on observed packet loss, not theoretical maximum throughput. Oversizing wastes kernel memory.

Impact on TCP Connections

These rmem_max and wmem_max settings also limit TCP socket buffers. Higher limits improve throughput on high-latency or high-bandwidth links. Use this formula to calculate optimal buffer size:

buffer_size_bytes = bandwidth_Mbps × latency_ms / 8

For example, a 100 Mbps link with 50 ms latency needs at minimum: (100 × 50) / 8 = 625 KB. Setting buffers much higher than needed only increases kernel memory consumption.

Monitoring and Verification

Monitor kernel memory usage to ensure buffer tuning doesn’t strain the system:

free -h
cat /proc/meminfo | grep -E 'MemAvailable|Buffers|Cached|Slab'

Verify packet loss with ethtool (if your NIC driver supports it):

sudo ethtool -S eth0 | grep -i drop

For application-specific verification, check application logs or use tcpdump to capture packets:

sudo tcpdump -i any 'udp port 5353' -c 1000 -w /tmp/dns.pcap

Then analyze for gaps in sequence numbers or missing datagrams.

Important Caveats

← All posts