Hi:
Sorry for not being able to respond in time due to some system problems at my end. Your ideas for the problem were really helpful. By means of a series of testing I was able to see the coalescing of small application writes into a single large segment due to wait for receipt of acks from the other end. Sure, Nagle is not the complete picture. Essentially whenever the wait time for acks from the receiver is increased due to network problems etc, TCP combines the segments. I also observed that occassionally the packet size was oscillating between full size and application write size in some test runs.
Thanks a ton for your comments. I appreciate it.
-- Vikas Paliwal
----- Original Message ----- From: "Nivedita Singhvi" <niv@xxxxxxxxxx>
To: "Vikas Paliwal" <vpaliwal@xxxxxxxxxxxxxxxxxx>
Cc: <linux-net@xxxxxxxxxxxxxxx>
Sent: Thursday, February 03, 2005 2:30 AM
Subject: Re: Disabling Nagle Algorithm
Vikas Paliwal wrote:Hi,
Is there a way so that the Nagle algorithm
can be disabled. I tried the TCP_NODELAY
socket option but that does not seem to help.
I need to send specific sized packets for an experimental evaluation. The TCP_NODELAY
option seems to work for a while, after which it switches
Nagle merely institutes a delay when you don't have at least half a segment - this allows the kernel to collect more data and send out a full-sized segment more often, avoiding sending out a stream of small packets. It is entirely possible that by the time the kernel gets an ack back from the receiver for the first sends (slow start) your app has written enough data for a full sized segment to go out - and Nagle would not come into play at all in that case. It all depends on how your application is written and what sized writes you are doing, etc..
to its size 1448 bytes. For example, I have an application
that writes packets of size much smaller than 1448 bytes.
Tcpdump shows that initially the packet size is the chosen small
value, but after a while it changes back to 1448 bytes (appears
like some buffer getting full). The time after which this change occurs also varies.
Can you tell us what is the sequence of sends (size) - tcpdump log would help.
It would be very helpful if somebody could provide some insight into this or we can conclude either of the two:
1. No socket or kernel option can guarantee disabling of Nagle algorithm and there is no way packet size can be chosen as in UDP.
Since TCP is a byte stream - you can only really control the segment sizes going out by having your application enforce that by waiting for data in the reverse direction or something. Turning off Nagle only helps, but is not the entire picture..
thanks,
Nivedita
- : send the line "unsubscribe linux-net" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
