[e2e] Traffic Burstiness Survey

[Neil Davies]

I've always found the coefficient of variance of the offered load to be a very useful indicator.


On 10 Sep 2012, at 07:02, Anoop Ghanwani <anoop at alumni.duke.edu> wrote:

> On Sun, Sep 9, 2012 at 3:39 PM, Craig Partridge <craig at aland.bbn.com> wrote:
>> 
>> I'll make a quick answer in the hopes that someone on the list will provide
>> a better one.
>> 
>>> 1) =91Bursty=92 is a word with no agreed meaning. How do you define a burst=
>>> y
>>> traffic?
>> 
>> I don't know of a widely agreed upon definition.  Generally when people say
>> "bursty" traffic they mean non-isochronous traffic -- namely traffic where
>> transmissions are not evenly spaced.
> 
> I agree with this definition.  Traffic gets to be bursty because of
> the "on/off" nature of sources.
> 
>>> 2) If you are involved with a data center, is your data center traffic
>>> bursty?
>>>    -- If yes,
>>>        -- Do you think that it will be useful to supress the burstiness in
>>> your traffic? (For example by pacing the traffic into shorter bursts)
>>>    -- If no:
>>>        -- Are you already supressing the burstiness? How?
>>>         -- Would you anticipate the traffic becoming burstier in the
>>> future?
> 
> While I'm not a data center operator, I think I can safely
> say that most would categorize their traffic as bursty.
> Do they regard it as a problem?  Yes and no.  Yes, because
> it impacts the amount of buffering needed in devices.
> No because it can largely be solved by overprovisioning
> the network and bandwidth is cheap.  Avoiding/mitigating
> bursts typically involves shaping and that is expensive and
> introduces additional latency.
> 
> In general, there are many data center applications --
> storage, compute (e.g. hadoop, map reduce), etc. -- and
> the traffic patterns for each of those would have to be analyzed
> separately.
> 
> Anoop