How Many TCP Connections Can One Server Really Handle? A Deep Dive

Why many people are confused about concurrency

In network development many developers still don’t fully understand a basic question: how many network connections can a single server support at most? This article explains the answer.

A chat about server‑side concurrency

“The TCP connection four‑tuple consists of source IP, source port, destination IP and destination port. Changing any element creates a different connection. For my Nginx, the port (80) and IP are fixed, so only source IP and source port vary. Theoretically Nginx can establish 2^32 (IP count) × 2^16 (port count) connections, a number in the hundreds of trillions.”

“Each opened file (in Linux everything is a file, including sockets) consumes memory. If a malicious process opens unlimited files, the server will crash. Therefore Linux limits the number of open file descriptors at system, user and process levels.”

System level: maximum number of open files, configurable via

fs.file-max

User level: maximum per‑user open files, configurable in

/etc/security/limits.conf

Process level: maximum per‑process open files, configurable via

fs.nr_open

“The receive buffer size can be viewed and configured with the sysctl command.”

<code>$ sysctl -a | grep rmem
net.ipv4.tcp_rmem = 4096 87380 8388608
net.core.rmem_default = 212992
net.core.rmem_max = 8388608</code>

“In tcp_rmem the first value is the minimum bytes allocated for a TCP connection’s receive buffer, default 4 KB, maximum up to 8 MB.”

“The send buffer size is controlled by net.ipv4.tcp_wmem .”

<code>$ sysctl -a | grep wmem
net.ipv4.tcp_wmem = 4096 65536 8388608
net.core.wmem_default = 212992
net.core.wmem_max = 8388608</code>

“In tcp_wmem the first value is the minimum send buffer size, also 4 KB by default.”

Achieving a million connections on the server

“To run the experiment we raised the limits at system, user and process levels to 1.1 million (target 1 million). This ensures other commands like ps and vi remain usable.”

<code>$ ss -n | grep ESTAB | wc -l
1000024</code>

The machine has 3.9 GB total memory, with the kernel slab using about 3.2 GB. Free memory and buffers together are only about 100 MB:

<code>$ cat /proc/meminfo
MemTotal:        3922956 kB
MemFree:           96652 kB
MemAvailable:       6448 kB
Buffers:           44396 kB
...
Slab:          3241244KB kB</code>

Using

slabtop

we see that the kernel objects

densty

,

flip

,

sock_inode_cache

and

TCP

each have around one million instances.

Conclusion

High concurrency is a key characteristic of internet backend services. After reading this article you should have a clear understanding of how many TCP connections a single server can support and how to tune the system to reach that limit.