Network Tuning

LogZilla documentation for Network Tuning

Networking performance (UDP)

Steps to test UDP ingest performance and adjust kernel parameters. Commands are examples; ports and values must be verified for the environment.

Testing UDP performance

  • Use a non-conflicting high port for testing (for example, 5514) to avoid disrupting the running syslog receiver and to avoid privileged-port requirements.

  • Start a UDP listener using netcat:

    bash
    netcat -u -p 5514 -l > /tmp/logs

    Alternatively (on systems where nc is used):

    bash
    nc -u -l 5514 > /tmp/logs

  • In another terminal, generate messages using loggen (part of syslog-ng):

    bash
    ./loggen -r 10000 -D -I 10 127.0.0.1 5514

  • Example output from loggen:

    text
    average rate = 10877.62 msg/sec, count=108783, time=10.006, msg size=256,
bandwidth=2719.40 kB/sec

  • Verify line count matches (or closely matches) the loggen reported count:

    bash
    wc -l /tmp/logs

    Example:

    text
    108783 /tmp/logs

  • Check UDP statistics:

    bash
    netstat -su

    Modern equivalent using ss:

    bash
    ss -u -s   # UDP summary
ss -u -a   # all UDP sockets

    Key fields:

    • packets received: total packets since last boot.
    • packets to unknown port: packets sent to a closed port (for example, service stopped while senders keep sending).
    • packet receive errors: errors while receiving/processing; a single packet can generate multiple errors.

Tuning UDP buffers

If netstat -su shows errors, increase UDP buffer sizes incrementally and re-test after each change.

  • Increase receive buffer max (example: 32 MB):

    bash
    sysctl -w net.core.rmem_max=33554432

    The default on many Linux distributions is about 122 KB (net.core.rmem_default = 124928).

  • Additional tuning options (examples):

    bash
    # UDP memory (in pages), min rmem
# NOTE: values are environment-dependent; increase gradually and test
echo 'net.ipv4.udp_mem = 192576 256768 385152' >> /etc/sysctl.conf
echo 'net.ipv4.udp_rmem_min = 4096' >> /etc/sysctl.conf

# Alternative runtime example for udp_mem
sysctl -w net.ipv4.udp_mem='262144 327680 393216'

    Optional transmit buffer tuning (device-dependent):

    bash
    sysctl -w net.core.rmem_default=212992
sysctl -w net.core.wmem_default=212992
sysctl -w net.core.wmem_max=33554432

    net.ipv4.udp_mem values are page-based. With PAGE_SIZE=4096, the maximum above equals 385152 * 4096 = 1,577,582,592 bytes.

  • Increase the incoming packet queue size:

    bash
    sysctl -w net.core.netdev_max_backlog=2000

Persisting settings

Settings changed with sysctl -w are not persistent. To make the changes permanent, append to /etc/sysctl.conf and reload:

bash
# Example: persist rmem_max at 32 MB
echo 'net.core.rmem_max=33554432' >> /etc/sysctl.conf

# Apply changes
sysctl -p

Re-run tests after each change to validate impact and adjust values as needed.

Alternatively, create a drop-in file and reload all settings:

bash
cat >/etc/sysctl.d/99-logzilla-tuning.conf <<'EOF'
net.core.rmem_max=33554432
EOF

sysctl --system

Advanced NIC tuning (optional)

These options can improve performance on multi-queue NICs in high-throughput environments. Hardware and kernel support vary; test changes incrementally in staging before applying to production.

Multi-queue visibility

bash
ls -d /sys/class/net/eth0/queues/rx-* 2>/dev/null || true
ls -d /sys/class/net/eth0/queues/tx-* 2>/dev/null || true

Receive Packet Steering (RPS) and Receive Flow Steering (RFS)

View and set global flow entries (example values):

bash
cat /proc/sys/net/core/rps_sock_flow_entries
echo 32768 > /proc/sys/net/core/rps_sock_flow_entries

Distribute per-queue flow counts and CPUs (example; adjust mask to available CPUs, for example f = CPUs 0–3):

bash
for q in /sys/class/net/eth0/queues/rx-*/rps_flow_cnt; do echo 4096 > "$q"; done
for q in /sys/class/net/eth0/queues/rx-*/rps_cpus; do echo f > "$q"; done

Transmit Packet Steering (XPS)

Pin transmit queues to CPUs (example; adjust CPU mask):

bash
for q in /sys/class/net/eth0/queues/tx-*/xps_cpus; do echo f > "$q"; done

Note: RPS/RFS/XPS settings reset on reboot. Persist via udev rules or a systemd unit executed at boot. Not all systems support xps_cpus, in which case it will indicate write error: No such file or directory.

NIC ring buffers

Inspect and (if supported) increase ring sizes cautiously:

bash
ethtool -g eth0          # query
ethtool -G eth0 rx 4096 tx 4096   # set (hardware/driver dependent)

NIC offload features

View current offload settings:

bash
ethtool -k eth0

Only disable specific offloads during targeted troubleshooting. Defaults are generally optimal.

Interrupt mapping visibility

bash
grep -E 'eth0' /proc/interrupts

Most distributions run irqbalance to spread IRQ load automatically. Manual pinning is advanced and environment-specific.

Network Tuning | LogZilla Documentation