The article says the same thing - turning on TCP_NODELAY improved latency and throughput on a certain workload.

Nagle's algorithm was introduced to reduce the overhead of TCP/IP packets (i.e. smaller a packet, bigger the overhead). It is basically to prevent a poor user application from generating too many small packets. For example:

socket.write(new byte[] { 1 });
socket.write(new byte[] { 2 });
socket.write(new byte[] { 3 });
socket.write(new byte[] { 4 });

which will result in 4 packets. With Nagle's algorithm, it is internally reduced to a single packet. Nagle's algorithm is turned on by default, but do we need this behavior by default? I don't think so because:

We already have a buffer and a user has to issue a flush operation.
A user can use the write() operation, but with the codec framework, it's unlikely that it will be inefficient.
The overhead of TCP/IP packet is not big anymore. We have much wider bandwidth than when Nagle's algorithm was introduced. With 40 byte overhead, at worst case, an application has to pump apx. 3 million packets per second to saturate a gigabit link.
We might get false positive bug report about low throughput.

Some good discussion about this: https://bugzilla.mozilla.org/show_bug.cgi?id=542401

Uh oh!

Consider turning on TCP_NODELAY by default #939

Description

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions