* Re: TCP Spike
2007-09-27 3:58 TCP Spike Majumder, Rajib
@ 2007-09-27 14:33 ` Stephen Hemminger
2007-09-27 19:34 ` Ilpo Järvinen
1 sibling, 0 replies; 3+ messages in thread
From: Stephen Hemminger @ 2007-09-27 14:33 UTC (permalink / raw)
To: Majumder, Rajib
Cc: 'netdev@vger.kernel.org',
'linux-kernel@vger.kernel.org'
On Thu, 27 Sep 2007 11:58:01 +0800
"Majumder, Rajib" <rajib.majumder@credit-suisse.com> wrote:
> Hi,
>
> We have observed 40ms latency spikes in TCP connections in "burst" type of traffic. This affects regular TCP sockets. We observed this issue in kernels of 2.4.21 and kernel 2.6.5.
Unfortunately, 2.6.5 is out of my short term memory at this point. I do remember that 2.6.5
used BIC for congestion control, and there were some math errors in the congestion control
logic that caused it to be way to aggressive.
>
> Aparently, this seems to be fixed in 2.6.19.
>
> Can someone throw some light on this?
My guess is that the addition of the SACK hinting might be the major win. The code
takes 3 passes over the SACK list, so with large outstanding data that was a major
bottleneck, not sure if it was 4ms worth though.
>
> Is this a congestion control/avoidance issue? What congestion control algorithm is used before 2.6.8?
Default congestion control in early 2.6 was BIC, then after CUBIC stabilized it was made the default in 2.6.19.
Another thing that may cause changes in latency is Appropriate Byte Counting (ABC).
It was added in 2.6.14, but then turned off by default in 2.6.18. The problem is
that ABC caused performance problems with some applications that sent messages
as many small writes.
--
Stephen Hemminger <shemminger@linux-foundation.org>
^ permalink raw reply [flat|nested] 3+ messages in thread
* Re: TCP Spike
2007-09-27 3:58 TCP Spike Majumder, Rajib
2007-09-27 14:33 ` Stephen Hemminger
@ 2007-09-27 19:34 ` Ilpo Järvinen
1 sibling, 0 replies; 3+ messages in thread
From: Ilpo Järvinen @ 2007-09-27 19:34 UTC (permalink / raw)
To: Majumder, Rajib
Cc: 'netdev@vger.kernel.org',
'linux-kernel@vger.kernel.org'
On Thu, 27 Sep 2007, Majumder, Rajib wrote:
> We have observed 40ms latency spikes in TCP connections in "burst" type of traffic.
> This affects regular TCP sockets.
Are segments being sent full-sized, or is there perhaps some Nagle
component in it as well? I.e., are the applications using TCP_NODELAY?
> We observed this issue in kernels of 2.4.21 and kernel 2.6.5.
>
> Aparently, this seems to be fixed in 2.6.19.
>
> Can someone throw some light on this?
I think somebody, probably Alexey, enabled sending of ACK on every 2nd
segment. Previously small segment senders playing with Nagle were
complaining every now and then about performance because two small
segments did not generate ACKs but one had to accumulate, IIRC, half MSS
worth of data before ACK was sent. Could this be related to your case?
...In case you're having too much time, you can always try bisecting it
which finds out the causing commit... :-)
> Is this a congestion control/avoidance issue?
Congestion control is basically ACK clocked math for cwnd, ssthresh, etc.
state, which then results in permission to send new segments out etc.
(except for RTO part of course, which I'll ignore in the next statement).
Any delay gaps to sent packet after ACK receival, which triggered the
state changing math, isn't there due to congestion control but due to
other factors! 40ms is much below MIN_RTO (200ms), so it shouldn't be due
to RTO either... Note that also delayed ACKs are exception to the general
rule.
Congestion control is controlled like your CPU is. In your CPU there's
this whatever GHz clock which determines when the state changing events
take place, state changes don't just happen arbitarily but are _clocked_
(ACK _clocked_ in case of congestion control). Of course there will be
some propagation delay after the change to put in effect all the state
changes that are result of what occurred at clock edge (and this delay
assimilating to processing delay in the context of congestion control).
--
i.
^ permalink raw reply [flat|nested] 3+ messages in thread