Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.nebius.com/llms.txt

Use this file to discover all available pages before exploring further.

TCP throughput depends on how much unacknowledged data a sender and receiver can keep in flight. On long-distance or high-latency connections, the default Linux TCP buffer limits can constrain a single flow before the network link is saturated. This article shows how to increase TCP receive and send window limits on a Linux virtual machine (VM) in Nebius AI Cloud. Use this tuning if your workload transfers large amounts of data over connections with both:
  • High bandwidth
  • Noticeable round-trip latency
Typical examples include:
  • Cross-region replication
  • Large model or dataset transfers
  • Long-lived TCP sessions that need higher per-flow throughput
If your traffic is short-lived, latency-sensitive, or limited by the application rather than the network path, this tuning may have little effect. The following settings increase the maximum TCP buffer sizes while keeping default behavior unchanged: 
net.ipv4.tcp_rmem = 4096 131072 268435456
net.ipv4.tcp_wmem = 4096 16384 268435456
net.core.rmem_max = 536870912
net.core.wmem_max = 536870912
net.ipv4.tcp_adv_win_scale = 1
These values keep the minimum and default TCP buffer settings close to the Linux defaults, while increasing the maximum values for high-bandwidth links. The recommended maximum TCP window targets a single flow of about 3 Gbit/s over a 300 ms round-trip-time path. That bandwidth-delay product is about 108 MB, which is rounded up to 128 MiB and then doubled to account for the effect of net.ipv4.tcp_adv_win_scale=1.
Larger TCP buffers can increase memory consumption under heavy connection fan-out. Verify available memory before applying this tuning on smaller VMs or on hosts with many simultaneous high-throughput flows.

​
How to tune TCP window sizes

​
Prerequisites

Make sure that:
  1. The VM runs Linux.
  2. You can connect to the VM by using SSH. For more information, see How to connect to virtual machines in Nebius AI Cloud.
  3. You have sudo privileges on the VM.

​
Apply the changes

  1. Connect to the VM: 
    ssh <username>@<VM_IP_address>
  2. Create a sysctl configuration file: 
    sudo tee /etc/sysctl.d/90-nebius-tcp-window.conf >/dev/null <<'EOF'
net.ipv4.tcp_rmem = 4096 131072 268435456
net.ipv4.tcp_wmem = 4096 16384 268435456
net.core.rmem_max = 536870912
net.core.wmem_max = 536870912
net.ipv4.tcp_adv_win_scale = 1
EOF
  3. Apply the configuration: 
    sudo sysctl --system

​
Verify the configuration

Check that the VM reports the configured values: 
sysctl \
  net.ipv4.tcp_rmem \
  net.ipv4.tcp_wmem \
  net.core.rmem_max \
  net.core.wmem_max \
  net.ipv4.tcp_adv_win_scale
Expected output: 
net.ipv4.tcp_rmem = 4096 131072 268435456
net.ipv4.tcp_wmem = 4096 16384 268435456
net.core.rmem_max = 536870912
net.core.wmem_max = 536870912
net.ipv4.tcp_adv_win_scale = 1
If your application still does not reach the expected throughput, check the end-to-end path for other limits such as application-level buffer sizing, rate limiting, packet loss or congestion on the remote side.

​
Roll back the changes

To remove the configuration: 
sudo rm /etc/sysctl.d/90-nebius-tcp-window.conf
sudo sysctl --system
After rollback, the VM uses the remaining sysctl configuration from the operating system image and any other files in /etc/sysctl.d.