When implementing network bonding with two 1Gbps adapters, many engineers expect to see a combined 2Gbps throughput. However, iperf often reports only ~930Mbps - essentially the speed of a single NIC. This occurs even with proper LACP (802.3ad) configuration on both servers and switches.

Different bonding modes yield different throughput characteristics:

# Common bonding modes and their typical throughput:
balance-rr (Round Robin)    → ~940Mbps (single flow)
802.3ad (LACP)             → ~520-940Mbps (asymmetric)
balance-xor                → ~940Mbps
active-backup              → ~940Mbps

The fundamental constraint comes from TCP's single-flow behavior. Even with bonded interfaces:

• Each TCP connection gets hashed to a single physical NIC
• iperf's default single-threaded test (-P 1) can't utilize multiple paths
• LACP load balancing works per-flow, not per-packet

To properly test bonded throughput, use parallel streams:

# Server:
iperf -s

# Client (10 parallel streams):
iperf -c server_ip -P 10 -t 60 -i 10

This should show aggregated throughput approaching 2Gbps if:

• Switch LACP configuration is correct
• Traffic uses multiple TCP/UDP flows
• No other bottlenecks exist (CPU, disk, etc.)

For Netgear GS728TS, verify these LACP settings:

# Sample configuration snippet:
configure lacp ports 1,2 actor-key 1
configure lacp ports 1,2 admin-key 1
enable lacp ports 1,2

Critical parameters to check:

• LAG hashing algorithm (should be layer2+3)
• Flow control settings (IEEE 802.3x enabled)
• Port speed/duplex auto-negotiation

To identify where packets are actually flowing:

# Check interface statistics:
cat /proc/net/bonding/bond0

# Real-time traffic monitoring:
iftop -i bond0
nload -m bond0

For deeper analysis:

# Capture traffic on physical interfaces:
tcpdump -i eth0 -w eth0.pcap
tcpdump -i eth1 -w eth1.pcap

Consider these tools for more comprehensive testing:

# Multi-threaded UDP testing:
iperf3 -c server_ip -u -b 2G -P 8

# Sockperf for latency measurement:
sockperf ping-pong -i server_ip --tcp

Here's a complete bonding configuration for Ubuntu:

# /etc/network/interfaces
auto bond0
iface bond0 inet static
    address 192.168.1.100
    netmask 255.255.255.0
    slaves eth0 eth1
    bond-mode 802.3ad
    bond-miimon 100
    bond-lacp-rate 1
    bond-xmit-hash-policy layer2+3

---

Linux bonding offers several modes with different characteristics:

# Common bonding modes:
modprobe bonding mode=0   # balance-rr (Round Robin)
modprobe bonding mode=4   # 802.3ad (LACP)
modprobe bonding mode=6   # balance-alb (Adaptive Load Balancing)

iperf measures single TCP stream performance by default. Since most bonding modes (including LACP) maintain flow consistency (same NIC for a given connection), you'll never exceed single NIC bandwidth for a single stream. This explains your 930-945Mbps results.

Your multi-connection test (-P 10) should theoretically aggregate bandwidth, but several factors can prevent this:

# Better multi-stream test command:
iperf3 -c 192.168.1.2 -P 8 -t 30 -O 3 -i 1

The RX/TX imbalance you observe in 802.3ad mode is actually expected behavior. LACP uses hashing algorithms (typically layer2+3 or layer3+4) to determine path selection:

# Check your hash policy:
cat /sys/class/net/bond0/bonding/xmit_hash_policy

Your Netgear switch needs proper LACP configuration:

# Sample LACP config for Cisco (concepts apply to Netgear):
interface Port-channel1
  description LACP to ServerA
  switchport mode trunk
  channel-protocol lacp
  channel-group 1 mode active

For comprehensive testing, consider these tools:

# Use nuttcp for parallel testing:
nuttcp -T 30 -u -R -w4m -N8 192.168.1.2

# Or use multiple iperf instances:
for i in {1..4}; do
  iperf3 -c 192.168.1.2 -p 520$i -t 20 &
done

Several factors affect aggregated throughput:

• PCIe bus saturation
• CPU affinity/interrupt balancing
• TCP window scaling
• NIC queue configuration

Check these settings for optimal performance:

# Check interrupt balancing:
cat /proc/interrupts | grep eth
# Set CPU affinity:
echo 2 > /proc/irq/24/smp_affinity