[Tsvwg] Re: [e2e] Are you interested in TOEs and related issues
zec at tel.fer.hr
Mon Mar 8 15:46:30 PST 2004
On Monday 08 March 2004 22:50, Craig Partridge wrote:
> In message
> <8052E2EA753D144EB906B7A7AA399714022A05B7 at mailserv.hatteras.com>, "S
> tephen Suryaputra" writes:
> >I think Marko is referring to the situation when there is a
> > congestion. Because of that the queue builds up and if the queue
> > space is in term of the number of packets, then a bunch of small
> > packets can potentially dominate the queue and left little room for
> > large packets eventhough there is a space in the buffer to
> > accomodate the large ones.
> Well, I tried to nail this down by setting it up as a queueing system
> but that's awfully hard. Here's the cut:
> 1. Basically you have a general arrival process delivering pieces of
> work of variable size, at arrival rate A subject to the condition
> that if there's an event of size w1 at time t1, then the next event
> cannot arrive sooner than time t1+w1. [I.e. A(t2,w2) - A(t1,w1) >=
> 2. You then have a series of queues Q1-Qn attached to servers S1-Sn
> that serve events at a fixed rate R, where R1 <= A (i.e. we can
> handle the maximum rate).
> 3. You then have a departure queue D1, and departure server DS,
> which has a service rate D, where D can be defined such that D(w) <=
> w (that is, the output is faster or equivalent speed to the input) or
> D(w) >= w (output is slower).
> If D(w) <= w, then I think one can argue simply by inspection that
> the total work unit buffering in the system is [max(w)/min(w)]-n.
> That is you simply need enough queueing to handle a queue that
> develops because a big packet causes a bunch of little packets to
> queue behind it (because they arrive while the big packet is going
> out). See Comment (**) below.
> I can't find an easy solution for D(w) >= w, because, in general, the
> queue can grow without bound.
Isn't the TCP's ultimate performance goal precisely the maintenance of
the equilibrium D(w) = w in the bottleneck queue? If you are a lucky
owner of just the TCP implementation that has no problem of achieving
that goal for the particular network configuration, the queue length on
the bottleneck link will fluctuate - sometimes it will be almost empty,
sometimes almost full, and every now and then it will even have to drop
a packet or two, but hopefully not more than that.
All I was trying to point out in my prior posts is that precisely in
such a situation bursty traffic will present a big failure. If the
burst arrives when the buffer is almost full, then the majority of the
packets from the burst will be discarded. Furthermore, if packets of
smaller size are used, requiring more packets to carry the same amount
of useful payload, the greater will be the percentage of such packets
to be dropped, assuming the buffer size on the router is defined by the
number of packets it can hold, regardless of packet size.
> ** If you believe this model, what it says is that if you go with
> packet size buffers, and the ratio between the smallest possible
> packet and largest possible packet is big, then you can find your
> queues growing quite large, due not to classic congestion (where
> there is too much demand fighting for one output link) but rather to
> serialization delays.
More information about the end2end-interest