I have performed the testing. I sent segments with size of 32 bytes in
a loop. NuttX source code is based on SHA-1 =
1e2560267898f413c93c5fb616ddc5b1d4d07184.
As a TCP host I used Linux kernel 4.19.
In case of forcing TCP_QUICKACK option to 1 in Linux I added the
following line after recv():
recv(fd, in_buffer, MAX_BUFFER_SIZE, 0);
setsockopt(fd, IPPROTO_TCP, TCP_QUICKACK, (int[]){1}, sizeof(int));
Forcing TCP_QUICKACK option to 1 in Linux means forced disabling
delayed ACKs.
In case of forcing TCP_QUICKACK option to 0 in Linux I added the
following line after recv():
recv(fd, in_buffer, MAX_BUFFER_SIZE, 0);
setsockopt(fd, IPPROTO_TCP, TCP_QUICKACK, (int[]){0}, sizeof(int));
Forcing TCP_QUICKACK option to 0 in Linux means forced enabling delayed
ACKs.
If TCP_QUICKACK option is not set (not forced to any value) after each
recv() call, Linux kernel implements its adaptive algorithm on-the-
flight.
The results are as follows:
- Exp.1 (CONFIG_NET_TCP_SPLIT is disabled in NuttX, Linux socket using
default settings): ACKs come immediately, no issues.
- Exp.2 (CONFIG_NET_TCP_SPLIT is disabled in NuttX, forcing
TCP_QUICKACK option to 1 in Linux socket): ACKs come immediately, no
issues.
- Exp.3 (CONFIG_NET_TCP_SPLIT is disabled in NuttX, forcing
TCP_QUICKACK option to 0 in Linux socket): ACKs are delayed by 40 ms,
performance issues.
- Exp.4 (CONFIG_NET_TCP_SPLIT is enabled in NuttX with default
NET_TCP_SPLIT_SIZE [=40], Linux socket using default settings): ACKs
come immediately, no issues, packet split does not occur.
- Exp.5 (CONFIG_NET_TCP_SPLIT is enabled in NuttX with default
NET_TCP_SPLIT_SIZE [=40], forcing TCP_QUICKACK option to 1 in Linux
socket): ACKs come immediately, no issues, packet split does not occur.
- Exp.6 (CONFIG_NET_TCP_SPLIT is enabled in NuttX with default
NET_TCP_SPLIT_SIZE [=40], forcing TCP_QUICKACK option to 0 in Linux
socket): ACKs are delayed by 40 ms, performance issues, packet split
does not occur.
- Exp.7 (CONFIG_NET_TCP_SPLIT is enabled in NuttX with
NET_TCP_SPLIT_SIZE = 20, Linux socket using default settings): ACKs
come immediately, no issues, packets are splitted.
- Exp.8 (CONFIG_NET_TCP_SPLIT is enabled in NuttX with
NET_TCP_SPLIT_SIZE = 20, forcing TCP_QUICKACK option to 1 in Linux
socket): ACKs come immediately, no issues, packets are splitted.
- Exp.9 (CONFIG_NET_TCP_SPLIT is enabled in NuttX with
NET_TCP_SPLIT_SIZE = 20, forcing TCP_QUICKACK option to 0 in Linux
socket): ACKs come immediately, no issues, packets are splitted.
Thus if Linux socket (Kernel 4.19) is used with default settings or
if TCP_QUICKACK socket option is forced to 1, it does not matter if
CONFIG_NET_TCP_SPLIT is enabled or not. There are not performance
issues.
However, if TCP_QUICKACK socket option is forced to 0 in Linux, there
are performance issues in NuttX TCP sender unless CONFIG_NET_TCP_SPLIT
option in NuttX is enabled.
On Fri, 2021-10-15 at 07:54 -0600, Gregory Nutt wrote:
> I think we should be testing with smaller, more typical user buffer
> sizes to verify the performance when the split is disabled or
> removed.
