Re: tc question about ingress bandwidth splitting

Grant Taylor <[email protected]> Thu, 26 Mar 2020 04:03:42 +0000
Newsgroups org.kernel.vger.lartc
Message-ID <[email protected]>
On 3/25/20 9:44 PM, Philip Prindeville wrote:
> If they’re both oversubscribed, then how does it get divvied up?

Production gets their 40 & 8
Guest gets their 10 & 2
There's nothing left over to divvy up further.

> I’ve been the target of a DDoS reflection attack and 99% of my 
> traffic was TCP RST’s and ICMP Unreachable… and that’s just 
> what was getting through… not what was being dropped upstream.

Oy vey!

> Sure.  And by applying controls inside the firewall, you affect the 
> perceived end-to-end properties as seen by the sender.  Which is 
> about the best you can hope for.

Yep.

Though I have wondered about a VPN to a VPS where I could control the 
bulk of what comes in on my wire.  Or at least apply some QoS on the end 
sending to my link.  ;-)

> More worried about Netflix, Hulu, and Disney+ which are all TCP-based. 
> All three (and possibly Amazon Prime, I don’t remember) use HTTP 
> byte-ranges, but reuse the same connection.  So one connection, 
> but bursty fetches…

Fair enough.  I do more consuming with them and less technical analysis.

> And if they both want to go over quota… I guess they can compete.

Nope.  They both get their SLA.  Since there's nothing left over, 
there's nothing to compete for.

> No, they were clearing it because they thought they were protecting 
> subscribers with not-up-to-date equipment from being confused by 
> seeing markings they didn’t know how to correctly interpret.

 >:-|

> Odd, considering that customer equipment often moves faster than 
> ISP or RBOC’s.  The whole 5 years I was at Cisco, several RBOC’s 
> were still running 12.0S (and insisting on continued support) even 
> as I was writing features for 12.4(T)… And they had only recently 
> migrated off 11.3 Mainline.

I can't say as I'm surprised.

> That’s typically a configuration issue, and not a question of not 
> having current software.

ACK

> Unfortunately most of the surveys on how widely deployed ECN marking 
> in transit networks is, is 12-19 years old.

I should inquire of colleagues at $WORK.  Though we tend to focus on QUIC.

> Yeah, I was going to work at Ubiquiti on the OS update until they 
> made a salary offer…

~chuckle~

> I’m using Supermicro pizza boxes (mostly SYS-5018D’s) that require 
> EFI support…

I deployed more than a few Supermicro boxen in my time.

> Sure.  Though on Gigabit interfaces, 50mbps is not statistically 
> significant even if I blocked it out…

I was thinking more about the other 950 Mbps that might not be available 
for Production <-> Guest transfers.  (Assuming you're reaching from Prod 
into Guest through a stateful firewall to copy files or the likes.)

> I remember people having the same complaint about Reno back in 
> the day…

That was a little bit before my time.

> Ever wake up and realize “I’m old”?  Well, my wife wakes up every 
> day and says to me, “You’re old.”  But not the same thing…

Yes on all accounts.



-- 
Grant. . . .
unix || die
smime.p7s (application/pkcs7-signature, 3.9 KB) - not displayed