Hi everyone,
I think most of you now Akamai and how they deliver 20% of total internet content traffic...
This looks like a good explanation: http://research.microsoft.com/~ratul/akamai.html
Has anyone tried to build a system similar to Akamai? I should be possible to build it, in a smaller way of course. Some modified named (bind) servers, Squid, etc.
From my point of view, Akamai does the right thing: Why try to have lots
of peerings, good transit connnections, etc. when you can serve the content directly out of the most popular ISP networks. They don't need their own network infrastructure for content delivery services (hosting).
Is here anyone interested in this topic? Anyone has time and interest to build, "research" and test a "small Akamai" hosting system?
My idea is in general: - 2-3 providers (one of you?) - each ISP "donates" 2-3 servers for the project (physical or virtual) - find a modified bind and squid or rebuild it to do this Akamai-like "DNS and url magic" - write a lot of shell scripts for monitoring, etc. - test the bunch of magic with our company sites :P
... but maybe I'm just crazy and you might simply ignore this post :-)
thanks, have a nice day best regards Marco
Marco Fretz wrote: [..]
... but maybe I'm just crazy and you might simply ignore this post :-)
Most people know *how* to do it (fail-over anycast presto), the economics, deploying it worldwide and getting a good solid customer base factor are other factors though.
Now if you have 3-10 large content providers and then set this up for them, that is a good start, then get some capital, set up a couple of nodes around the world, and then more and more and more ;)
Greets, Jeroen
Jeroen Massar wrote:
Marco Fretz wrote: [..]
... but maybe I'm just crazy and you might simply ignore this post :-)
Most people know *how* to do it (fail-over anycast presto), the economics, deploying it worldwide and getting a good solid customer base factor are other factors though.
as I said, maybe it's a stupid idea. I'm more interested in the technical than in the economic aspect.
I've never set up IPv4 anycast services but as far as I know anycast is only recommended for connection less / session less services (DNS, etc...). I never saw http content delivered over anycast, because it's TCP and you need consistent data. pls correct me if I'm wrong...
greets Marco
Hi,
There is already some software there :
Looks quite dead, but seemed to work quite well ... Might be a start.
Le mardi 02 septembre 2008 à 10:38 +0200, Marco Fretz a écrit :
Hi everyone,
I think most of you now Akamai and how they deliver 20% of total internet content traffic...
This looks like a good explanation: http://research.microsoft.com/~ratul/akamai.html
Has anyone tried to build a system similar to Akamai? I should be possible to build it, in a smaller way of course. Some modified named (bind) servers, Squid, etc.
From my point of view, Akamai does the right thing: Why try to have lots of peerings, good transit connnections, etc. when you can serve the content directly out of the most popular ISP networks. They don't need their own network infrastructure for content delivery services (hosting).
Is here anyone interested in this topic? Anyone has time and interest to build, "research" and test a "small Akamai" hosting system?
My idea is in general:
- 2-3 providers (one of you?)
- each ISP "donates" 2-3 servers for the project (physical or virtual)
- find a modified bind and squid or rebuild it to do this Akamai-like
"DNS and url magic"
- write a lot of shell scripts for monitoring, etc.
- test the bunch of magic with our company sites :P
... but maybe I'm just crazy and you might simply ignore this post :-)
thanks, have a nice day best regards Marco
swinog mailing list swinog@lists.swinog.ch http://lists.swinog.ch/cgi-bin/mailman/listinfo/swinog
On Tue, Sep 02, 2008 at 11:01:23AM +0200, Marco Fretz wrote:
Jeroen Massar wrote:
Marco Fretz wrote: [..]
... but maybe I'm just crazy and you might simply ignore this post :-)
Most people know *how* to do it (fail-over anycast presto), the economics, deploying it worldwide and getting a good solid customer base factor are other factors though.
as I said, maybe it's a stupid idea. I'm more interested in the technical than in the economic aspect.
I've never set up IPv4 anycast services but as far as I know anycast is only recommended for connection less / session less services (DNS, etc...). I never saw http content delivered over anycast, because it's TCP and you need consistent data. pls correct me if I'm wrong...
Corrected, your wrong. TCP works just fine for short living TCP sessions (like 99% of all traffic). If you're routing is stable you always end up at the same site. Only on bgp route changes that influence the path to the anycast network you may get session drops because the traffic is flowing to a new site.
well, I am actually building one for a american company. End-Stage will be ~600 Processing servers (backend) and around 100 delivery Servers. I work with Squid.
contact me offlist if interested because I am partially under NDA (fredy is the only one yet that I have spoken about it with him as he is the guy to host it :) )
Silvan
Am 02.09.2008 um 11:23 schrieb Marcel Prisi:
Hi,
There is already some software there :
Looks quite dead, but seemed to work quite well ... Might be a start.
Le mardi 02 septembre 2008 à 10:38 +0200, Marco Fretz a écrit :
Hi everyone,
I think most of you now Akamai and how they deliver 20% of total internet content traffic...
This looks like a good explanation: http://research.microsoft.com/~ratul/akamai.html
Has anyone tried to build a system similar to Akamai? I should be possible to build it, in a smaller way of course. Some modified named (bind) servers, Squid, etc.
From my point of view, Akamai does the right thing: Why try to have lots of peerings, good transit connnections, etc. when you can serve the content directly out of the most popular ISP networks. They don't need their own network infrastructure for content delivery services (hosting).
Is here anyone interested in this topic? Anyone has time and interest to build, "research" and test a "small Akamai" hosting system?
My idea is in general:
- 2-3 providers (one of you?)
- each ISP "donates" 2-3 servers for the project (physical or
virtual)
- find a modified bind and squid or rebuild it to do this Akamai-like
"DNS and url magic"
- write a lot of shell scripts for monitoring, etc.
- test the bunch of magic with our company sites :P
... but maybe I'm just crazy and you might simply ignore this post :-)
thanks, have a nice day best regards Marco
swinog mailing list swinog@lists.swinog.ch http://lists.swinog.ch/cgi-bin/mailman/listinfo/swinog
swinog mailing list swinog@lists.swinog.ch http://lists.swinog.ch/cgi-bin/mailman/listinfo/swinog
Marco Fretz wrote:
Jeroen Massar wrote:
Marco Fretz wrote: [..]
... but maybe I'm just crazy and you might simply ignore this post :-)
Most people know *how* to do it (fail-over anycast presto), the economics, deploying it worldwide and getting a good solid customer base factor are other factors though.
as I said, maybe it's a stupid idea. I'm more interested in the technical than in the economic aspect.
Get a book/google a bit, take a box, setup a couple of XEN's on it, add some quagga's on them, presto, play with it. You just need some glue between the application you are running and quagga to tag if it is 'alive', simply killing bgpd/ospfd is good enough there.
One used to need a lot of boxes and networking hardware for these kind of tests, but long live VM's nowadays :)
Of course it depends on what kind of setup you want and what you actually want to achieve with it; but it ain't rocketscience, far from.
I've never set up IPv4 anycast services but as far as I know anycast is only recommended for connection less / session less services (DNS, etc...). I never saw http content delivered over anycast, because it's TCP and you need consistent data. pls correct me if I'm wrong...
The question you have to ask here is how often do you swap your anycast route? If that route is in flux, indeed, the packets will end up going to different servers and that won't work for TCP because of state; but if you have a stable route to your 'anycast' node, then all is fine. HTTP connections etc are generally quite short lived, eg say 10 seconds, max, thus if you anycast route swaps every 15 minutes, you are quite fine; the connections that break won't be really noticed as they will hopefully restart. Most Web-applications won't notice the failover.
Note that the actual anycast-node (the IP you are talking to) generally is a big fat loadbalancer which does know about TCP state (and actually terminates the TCP connection in some cases) and just forwards it on to a backend box which does the real work.
Enjoy your playtime ;)
Greets, Jeroen
Claudio Jeker wrote:
Corrected, your wrong. TCP works just fine for short living TCP sessions (like 99% of all traffic). If you're routing is stable you always end up at the same site. Only on bgp route changes that influence the path to the anycast network you may get session drops because the traffic is flowing to a new site.
Ok, thank you Jeroen and Claudio for this explanations about anycasting. But is there anyone using anycasts for HTTP content? I think its only used for DNS, etc... am I wrong?
Marco Fretz wrote:
Claudio Jeker wrote:
Corrected, your wrong. TCP works just fine for short living TCP sessions (like 99% of all traffic). If you're routing is stable you always end up at the same site. Only on bgp route changes that influence the path to the anycast network you may get session drops because the traffic is flowing to a new site.
Ok, thank you Jeroen and Claudio for this explanations about anycasting. But is there anyone using anycasts for HTTP content? I think its only used for DNS, etc... am I wrong?
Wrong. google(anycast http) google(anycast) google(distributed content system) etc etc etc etc etc :)
Greets, Jeroen
Jeroen Massar wrote:
[..] Ok, thank you Jeroen and Claudio for this explanations about anycasting. But is there anyone using anycasts for HTTP content? I think its only used for DNS, etc... am I wrong?
Wrong. google(anycast http) google(anycast) google(distributed content system) etc etc etc etc etc :)
oh, hmm.. ok :-) sorry. Now I feel like a noob ^^
Greets, Jeroen
On the Tue, Sep 02, 2008 at 10:38:08AM +0200, Marco Fretz blubbered:
Hallo.
I think most of you now Akamai and how they deliver 20% of total internet content traffic...
This looks like a good explanation: http://research.microsoft.com/~ratul/akamai.html
Has anyone tried to build a system similar to Akamai? I should be possible to build it, in a smaller way of course. Some modified named (bind) servers, Squid, etc.
On a side note, there is the Coral Content Distribution Network, that can be used for free for single files or websites for free. It acts like a giant worldwide proxy as I understand it.
CU, Venty
-
Claudio Jeker -
Jeroen Massar -
Marcel Prisi -
Marco Fretz -
Martin Ebnoether -
Silvan Gebhardt