[GRLUG] Testing Comcast vs AT&T

[Adam Tauno Williams]

On Mon, 2011-07-11 at 09:18 -0400, Michael Mol wrote:
> Then call it "meaningless in a general context"

Which generally translates to the shorter form: "meaningless".  :)

> Yes, the test has meaning for particular contexts. When looking
> specifically at one's internet connection, it only means "I know my
> internet connection *can* go at *least* this fast, for this kind of
> transferred." It doesn't speak to a broader-term connection maximum,
> or a broader-term connection minimum.

+1

> There are several parts of the problem:
> 1) People use tests like speedtest as a reflection of their internet
> connection, to the ignorance of other potential network effects. If I
> had run that speed test over 802.11b, for example, my throughput would
> have been far worse. If I had run that speed test on a day when
> comcast's edge routers were having difficulties (say there were a
> failing NIC somewhere, or backhoe fade mid-transfer), my throughput
> would have been far worse.

Which emphasizes the need use l-o-n-g lived tests (which are actually not
tests, just measurements).

> 2) People expect an internet connection like cable, ADSL or VDSL to
> have precise, measurable behavior. It doesn't, it won't and it can't.

+1

> Put another way, I'm paying $80/mo for just cable internet. By my
> speed tests, I'm getting anywhere from 42% to 82% of the throughput of
> a T3. At something like 10% of the cost*. And people like Bob think I
> should be raising hell because I'm not getting 44% instead of 42%, or
> not getting 82% all the time. That's just nuts.

It is also important to measure the actually level of traffic, rather
than the volume of "data" moved.  Because network protocols have
overhead; so the *only way* (for emphasis, I'll repeat "**only [valid]
way*") way to measure is at the *interface* and not at the *result* of
the conversation.

Since speedtest.net and other gimmicks can't do that... they are
bogus.  

IPv4, depending upon the use-case, can have horribly low efficiency;
especially if the nagle implementation is bad.  With SSH / telnet, for
example, IPv4 is as much as 89% overhead [IPv6 or SCTP over IPv4 can
help this quite a bit].  But IPv4 is 52 bytes of header for every packet
and Ethernet adds another 38 bytes per frame.  And then some more if you
have 802.1q.  And then this is 802.1x authentication...

> * Ok, I don't know what a T3 costs. Last time I looked hard at those
> numbers, a *T1* cost a couple grand a month, 

Nah, a T1 now costs ~$300 - $800 depending on where you are and how many
other circuits you can bundle into the contract.




-- 
This message has been scanned for viruses and
dangerous content by MailScanner, and is
believed to be clean.