Skip site navigation (1)Skip section navigation (2)
Date:      Fri, 30 Mar 2007 03:14:32 -0700
From:      "'Luigi Rizzo'" <rizzo@icir.org>
To:        Dave Raven <dave@raven.za.net>
Cc:        freebsd-ipfw@freebsd.org
Subject:   Re: Using "delay" to emulate a satellite link
Message-ID:  <20070330031432.A78468@xorpc.icir.org>
In-Reply-To: <01be01c772b2$91e64180$b5b2c480$@za.net>; from dave@raven.za.net on Fri, Mar 30, 2007 at 12:02:28PM %2B0200
References:  <015d01c77297$953ba250$bfb2e6f0$@za.net> <20070330005916.A76128@xorpc.icir.org> <01a501c772ab$1dcbd460$59637d20$@za.net> <20070330021442.A77652@xorpc.icir.org> <01be01c772b2$91e64180$b5b2c480$@za.net>

next in thread | previous in thread | raw e-mail | index | archive | help
On Fri, Mar 30, 2007 at 12:02:28PM +0200, Dave Raven wrote:
> Hi Luigi, 
> 	Firstly sorry for the confusion - I mean even if I change it to
> 250/250 (e.g. 500ms of delay) I can't reach 1mbps of throughput. Is it a
> definite that because of the latency (if its bad enough) you can never
> achieve full throughput - because of TCP ack delays?
> 
> Your 65kbytes/1000ms = 512kbit/s calculation - is that 65k from the
> send/recv space ? The maximum is 65k correct ? Does that mean that on a
> given 1000ms link you CANNOT achieve more than 512kbps ?

yes this is basic networking stuff - for a window-based protocol
the max throughtput is 1 window per rtt, where the window is
upper bounded by the min of socket buffer, tcp buffers, negotiated
tcp window

luigi


> Thanks so much for the help - I know its going a bit off topic
> 
> Dave
> 
> -----Original Message-----
> From: 'Luigi Rizzo' [mailto:rizzo@icir.org] 
> Sent: Friday, March 30, 2007 11:15 AM
> To: Dave Raven
> Cc: freebsd-ipfw@freebsd.org
> Subject: Re: Using "delay" to emulate a satellite link
> 
> On Fri, Mar 30, 2007 at 11:09:07AM +0200, Dave Raven wrote:
> > Hi Luigi,
> > 	Thanks for the reply -- these are my related send/recv space
> > settings on all of the boxes involved --
> > 
> > net.local.stream.sendspace: 65535
> > net.local.stream.recvspace: 65535
> > net.local.dgram.recvspace: 4096
> > net.inet.tcp.sendspace: 65535
> > net.inet.tcp.recvspace: 65535
> > net.inet.udp.recvspace: 65535
> > net.inet.raw.recvspace: 8192
> > 
> > What I have done is put 3 freebsd units plugged into eath other with
> > crossover cables, like so --
> > 
> > BOX 1 ----- BOX 2 ----- BOX 3
> > 
> > Box 3 has a 500meg /dev/urandom text file on a webserver. Box 1 wget's
> that
> > file. Box 2 has the ipfw delay setup. If I ping from BOX 1 to BOX 3 I get
> > +/-1000ms of delay. When I download I get less than half of the link speed
> > (1meg limited by box 2).
> 
> which is fine - you get 65kbytes/1000ms = 512kbit/s
> 
> > The latency directly affects the throughput - but my question is this --
> If
> > I download something from a web server on the internet that I have 500ms
> of
> > delay to, I can get 1meg (on a 1meg link). When I emulate that delay with
> > dummynet I can't -- is there a difference in the type of delay
> experienced?
> 
> well, it's twice as much in your case - your ping time is 1000ms instead
> of 500ms
> 
> cheers
> luigi
> 
> > Really anyone can - most people have delay of at least 300-500ms to remote
> > webservers (e.g. from me in Africa to America); but it doesn't hamper
> > download speed?
> > 
> > Thanks again
> > Dave 
> > 
> > 
> > 
> > -----Original Message-----
> > From: Luigi Rizzo [mailto:rizzo@icir.org] 
> > Sent: Friday, March 30, 2007 9:59 AM
> > To: Dave Raven
> > Cc: freebsd-ipfw@freebsd.org
> > Subject: Re: Using "delay" to emulate a satellite link
> > 
> > On Fri, Mar 30, 2007 at 08:49:19AM +0200, Dave Raven wrote:
> > > Hi all,
> > > 	I've been looking at the ipfw (dummynet) ability to do delay and
> > > have a few questions - I hope this is the right list. I want to simulate
> a
> > > 1000ms RTT on a satellite link. To do that I've created an inbound and
> > > outbound pipe and given each 1mb and 500ms of delay.
> > > 
> > > However, I'm unable to get anywhere near 1mb of throughput on it until I
> > > drop the delay. I believe I understand the slowdown due to the latency,
> > but
> > > my question is this - an http download through a 500ms "emulated" link
> > > that's running 1 mb can't get 1mb, yet if I download over the internet
> on
> > a
> > > site that's pinging 500ms, it goes 1mb 
> > > 
> > > Whats the difference between dummynet delay and real life distance
> delay?
> > 
> > first make sure you are not comparing apples and oranges.
> > what sender and receiver are you using to get 1Mbit/s on a 500ms link ?
> > and, are you sure that if you ping from the source to the destination
> > you are using for your tests with dummynet you get the delay
> > you are expecting (1000ms as you configured it, and not 2000 ?)
> > 
> > 500ms of delay on each pipe give at least 1000ms total
> > delay (assuming you have not misconfigured your dummynet box).
> > In order to fill the pipe you need at least 1Mbit worth of data
> > in the socket buffer/tcp window - the default on FreeBSD is
> > 32kbytes sending, 64Kbytes receiving, so you won't be
> > able to achieve that unless you increase these two sysctls:
> > 
> > 	net.inet.tcp.sendspace: 32768
> > 	net.inet.tcp.recvspace: 65536
> > 
> > if you draw the bw vs delay that you achieve on your connection
> > you will likely find that either the limit is your socket
> > buffers or a misconfigured ipfw which results in twice the delay
> > 
> > 	cheers
> > 	luigi



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20070330031432.A78468>