How to accelerate tcp retransmition

netdev.vger.kernel.org archive mirror
 help / color / mirror / Atom feed

* How to accelerate tcp retransmition
@ 2012-01-24  8:58 leno soff
  2012-01-24 14:09 ` leno soff
  0 siblings, 1 reply; 3+ messages in thread
From: leno soff @ 2012-01-24  8:58 UTC (permalink / raw)
  To: netdev; +Cc: linux-kernel, eric.dumazet, davem

Hello everyone,
   I have a lot of client, say C1, C2, .. Cn, which route packets to
my servers through router R. But R will drop 70% - 90% _ non-SYN _
packets from clients to Server.

  The clients and server work like http client and server, sending a
request, wating for the response.

  When router R drop the request-like packets(without SYN flag, of
cause), the clients retransmit with icsk_backoff++;  They usually
should wait 15 seconds+ (1+2+4+8), and sometimes more or just timeout.

   So, my question is:

        Is there any way in _server_ side, to accelerate the
retransmition like accelerating slow-start. I tried SO_KEEPALIVE in
server, interval == 1 second or less,  I assume the keepalive ack will
 accelerate the retransmition, ,  but it does not work:

       Flags [P.], seq 1:314, ack 1.  ----------------client request,
which is lost
       Flags [.], ack 1, win 905, length 0  --- server keepalive ack
       Flags [.], ack 1, win 115, length 0  --- client responds ack
       .... A lot keepalive and response.

    The client timeout because of RTO, why doesn't retransmit the
request packet but just an ack without data when it sees keepalive ack
?

Thanks.

^ permalink raw reply	[flat|nested] 3+ messages in thread

* Re: How to accelerate tcp retransmition
  2012-01-24  8:58 How to accelerate tcp retransmition leno soff
@ 2012-01-24 14:09 ` leno soff
  2012-01-24 17:57   ` Rick Jones
  0 siblings, 1 reply; 3+ messages in thread
From: leno soff @ 2012-01-24 14:09 UTC (permalink / raw)
  To: netdev; +Cc: linux-kernel, eric.dumazet, davem

On Tue, Jan 24, 2012 at 4:58 PM, leno soff <lnxuff@gmail.com> wrote:
> Hello everyone,
>   I have a lot of client, say C1, C2, .. Cn, which route packets to
> my servers through router R. But R will drop 70% - 90% _ non-SYN _
> packets from clients to Server.
>
>  The clients and server work like http client and server, sending a
> request, wating for the response.
>
>  When router R drop the request-like packets(without SYN flag, of
> cause), the clients retransmit with icsk_backoff++;  They usually
> should wait 15 seconds+ (1+2+4+8), and sometimes more or just timeout.
>
>   So, my question is:
>
>        Is there any way in _server_ side, to accelerate the
> retransmition like accelerating slow-start. I tried SO_KEEPALIVE in
> server, interval == 1 second or less,  I assume the keepalive ack will
>  accelerate the retransmition, ,  but it does not work:
>
>       Flags [P.], seq 1:314, ack 1.  ----------------client request,
> which is lost
>       Flags [.], ack 1, win 905, length 0  --- server keepalive ack
>       Flags [.], ack 1, win 115, length 0  --- client responds ack
>       .... A lot keepalive and response.
>
>    The client timeout because of RTO, why doesn't retransmit the
> request packet but just an ack without data when it sees keepalive ack
> ?
>
>
> Thanks.


After 3-way shakehands,  Client sends 9 bytes(I make sure it is
DROPPED by router),
I modify the server-side kernel, set keepalive timer interval = 100ms
to accelerate retransmiting 3+ DUP ACKS which makes sure the client
will not timeout.
I run 'ss -en|grep 8888' in client side, it shows the retransmition
timer exp-increasing which is not expected for me. I assume it will
retransimit after receiving 3+ DUP ACKS.
And below is the packets flow.
I think it is not the right way to solve this Bad-Router problem, but
it is the only method I know which failed.

In one word, I don't want the client timeout.

Any  suggestion is helpful
Thanks


21:41:17.586310 IP 192.168.10.22.49538 > 199.192.156.23.8888: Flags
[.], ack 1, win 115, options [nop,nop,TS val 38962812 ecr 12647162],
length 0
21:41:17.680113 IP 192.168.10.22.49538 > 199.192.156.23.8888: Flags
[P.], seq 1:9, ack 1, win 115, options [nop,nop,TS val 38962906 ecr
12647162], length 8
21:41:17.687612 IP 199.192.156.23.8888 > 192.168.10.22.49538: Flags
[.], ack 1, win 905, options [nop,nop,TS val 12673575 ecr 38962367],
length 0
21:41:17.687635 IP 192.168.10.22.49538 > 199.192.156.23.8888: Flags
[.], ack 1, win 115, options [nop,nop,TS val 38962913 ecr 12647162],
length 0
 --------- a lot keepalive here interval=100ms ------
21:41:18.418845 IP 199.192.156.23.8888 > 192.168.10.22.49538: Flags
[.], ack 1, win 905, options [nop,nop,TS val 12674313 ecr 38962367],
length 0
21:41:18.418862 IP 192.168.10.22.49538 > 199.192.156.23.8888: Flags
[.], ack 1, win 115, options [nop,nop,TS val 38963644 ecr 12647162],
length 0
21:41:18.427900 IP 192.168.10.22.49538 > 199.192.156.23.8888: Flags
[P.], seq 1:9, ack 1, win 115, options [nop,nop,TS val 38963654 ecr
12647162], length 8
21:41:18.520310 IP 199.192.156.23.8888 > 192.168.10.22.49538: Flags
[.], ack 1, win 905, options [nop,nop,TS val 12674413 ecr 38962367],
length 0
21:41:18.520332 IP 192.168.10.22.49538 > 199.192.156.23.8888: Flags
[.], ack 1, win 115, options [nop,nop,TS val 38963746 ecr 12647162],
length 0
-----------  a lot keepalive here too ----------
21:41:19.877055 IP 199.192.156.23.8888 > 192.168.10.22.49538: Flags
[.], ack 1, win 905, options [nop,nop,TS val 12675771 ecr 38962367],
length 0
21:41:19.877082 IP 192.168.10.22.49538 > 199.192.156.23.8888: Flags
[.], ack 1, win 115, options [nop,nop,TS val 38965103 ecr 12647162],
length 0
21:41:19.925886 IP 192.168.10.22.49538 > 199.192.156.23.8888: Flags
[P.], seq 1:9, ack 1, win 115, options [nop,nop,TS val 38965152 ecr
12647162], length 8
21:41:19.975419 IP 199.192.156.23.8888 > 192.168.10.22.49538: Flags
[.], ack 1, win 905, options [nop,nop,TS val 12675871 ecr 38962367],
length 0
21:41:19.975446 IP 192.168.10.22.49538 > 199.192.156.23.8888: Flags
[.], ack 1, win 115, options [nop,nop,TS val 38965201 ecr 12647162],
length 0

^ permalink raw reply	[flat|nested] 3+ messages in thread

* Re: How to accelerate tcp retransmition
  2012-01-24 14:09 ` leno soff
@ 2012-01-24 17:57   ` Rick Jones
  0 siblings, 0 replies; 3+ messages in thread
From: Rick Jones @ 2012-01-24 17:57 UTC (permalink / raw)
  To: leno soff; +Cc: netdev, linux-kernel, eric.dumazet, davem

On 01/24/2012 06:09 AM, leno soff wrote:
> On Tue, Jan 24, 2012 at 4:58 PM, leno soff<lnxuff@gmail.com>  wrote:
>> Hello everyone,
>>    I have a lot of client, say C1, C2, .. Cn, which route packets to
>> my servers through router R. But R will drop 70% - 90% _ non-SYN _
>> packets from clients to Server.

I think you should address that first - why does the router drop 70 to 
905 of the non-SYN segments it receives.  At first blush that sounds 
like a classic case of massive over-subscription, and any attempt to 
accelerate the retransmissions seems destined to trigger congestive 
collapse.

As a side issue, if the router is dropping 70 to 90% of the non-SYN 
segments, why is it not also dropping 70 to 90% of the SYN segments?

Also, I'm reasonably confident that an actual keepalive probe would not 
(or at least should not?) be considered in the three duplicate ACK 
heuristic.

rick jones

^ permalink raw reply	[flat|nested] 3+ messages in thread

end of thread, other threads:[~2012-01-24 17:57 UTC | newest]

Thread overview: 3+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2012-01-24  8:58 How to accelerate tcp retransmition leno soff
2012-01-24 14:09 ` leno soff
2012-01-24 17:57   ` Rick Jones

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).