s*****g
发帖数: 1055 | 1 Assuming we have a bunch of data centers across different geo-locations, the
connection among those data centers are over public Internet, applications
require that data centers can talk to each other with minimum latency,
applications itself will take care of data encryption so no VPN devices are
needed. Since we have no control of the underlying transport, we can have
the following situation: there are packet drops between Tokyo and Seattle
due to some transport providers link congestion, but the paths of Tokyo <-->
LA, LA<-->Seattle are good, unfortunately we have no ability to route Tokyo
-> Seattle traffic via LA. One solution I am thinking of is to build a full
-mesh GRE tunnels among the data centers and run IGP over GRE tunnels, so I
can have some level of control of the path. The big problem of doing this is
, how do I feed the underlying network performace back and dynamically
adjust the GRE metrics? the considered parameters can be latency, packet
drop rate, throughput etc.
thoughts? I am sure many other people faced or are
facing the same problem. |
c*****i
发帖数: 631 | 2 纯ip挺难的吧。你这个问题是SDN的经典应用之一，可惜现在各家的SDN都还在beta阶段
。如果是mpls te和isis倒是可以在cspf上看看。 |
s*****g
发帖数: 1055 | 3 MPLS TE won't help, first, We don't have our own transport, second, even if
we do, MPLS TE signaling is not aware of underlying physical link
utilization.
Theatrically I don't think there is any difficulty of implementing what I
want, I just don't have the expertise to write the code.
【在 c*****i 的大作中提到】
: 纯ip挺难的吧。你这个问题是SDN的经典应用之一，可惜现在各家的SDN都还在beta阶段
: 。如果是mpls te和isis倒是可以在cspf上看看。
|
z**r
发帖数: 17771 | 4 can wan/dns load balancing resolve this?
the
applications
are
->
Tokyo
full
I
【在 s*****g 的大作中提到】
: Assuming we have a bunch of data centers across different geo-locations, the
: connection among those data centers are over public Internet, applications
: require that data centers can talk to each other with minimum latency,
: applications itself will take care of data encryption so no VPN devices are
: needed. Since we have no control of the underlying transport, we can have
: the following situation: there are packet drops between Tokyo and Seattle
: due to some transport providers link congestion, but the paths of Tokyo <-->
: LA, LA<-->Seattle are good, unfortunately we have no ability to route Tokyo
: -> Seattle traffic via LA. One solution I am thinking of is to build a full
: -mesh GRE tunnels among the data centers and run IGP over GRE tunnels, so I
|
s*****g
发帖数: 1055 | 5 How?
【在 z**r 的大作中提到】
: can wan/dns load balancing resolve this?
:
: the
: applications
: are
: ->
: Tokyo
: full
: I
|
c*****i
发帖数: 631 | 6 你可以用ipsla+eem，但是我觉得不flexible和scalable。因为每个router做自己的没
有全局观，可能会出问题。SDN是最好的，但是现在还是beta，而且需要你对编程有相
当的要求。
if
【在 s*****g 的大作中提到】
: MPLS TE won't help, first, We don't have our own transport, second, even if
: we do, MPLS TE signaling is not aware of underlying physical link
: utilization.
: Theatrically I don't think there is any difficulty of implementing what I
: want, I just don't have the expertise to write the code.
|
s*****g
发帖数: 1055 | 7 Not every vendor supports ipsla monitoring, I am actually thinking of doing
this on Linux.
SDN is far from production, and the way I understand SDN, the controller has
to be close to the box so you can use OOB network for flow configuration, I
think Google is doing some kind of SDN but with off-line computation?
【在 c*****i 的大作中提到】
: 你可以用ipsla+eem，但是我觉得不flexible和scalable。因为每个router做自己的没
: 有全局观，可能会出问题。SDN是最好的，但是现在还是beta，而且需要你对编程有相
: 当的要求。
:
: if
|
a***n
发帖数: 262 | 8 So each site is multi-homed to the same or different ISPs?
In either case you won't have control
over the ISP's internal routing since you mentioned over public internet
and you might have control over the exit/entry point from/to you?
Cisco PfR if Cisco?
http://www.cisco.com/en/US/products/ps8787/products_ios_protoco
or put certain prefixes in different VRF on each site to measure the
performance of different path or PBR, then adjust your internal routing in
Linux?
doing
has
I
【在 s*****g 的大作中提到】
: Not every vendor supports ipsla monitoring, I am actually thinking of doing
: this on Linux.
: SDN is far from production, and the way I understand SDN, the controller has
: to be close to the box so you can use OOB network for flow configuration, I
: think Google is doing some kind of SDN but with off-line computation?
|
s*****g
发帖数: 1055 | 9 No, Cisco PrR/OER won't help in this situation either.
【在 a***n 的大作中提到】
: So each site is multi-homed to the same or different ISPs?
: In either case you won't have control
: over the ISP's internal routing since you mentioned over public internet
: and you might have control over the exit/entry point from/to you?
: Cisco PfR if Cisco?
: http://www.cisco.com/en/US/products/ps8787/products_ios_protoco
: or put certain prefixes in different VRF on each site to measure the
: performance of different path or PBR, then adjust your internal routing in
: Linux?
:
|
a***n
发帖数: 262 | 10 Borrow some idea from peer to peer protocols?
No, Cisco PrR/OER won't help in this situation either.
【在 s*****g 的大作中提到】
: No, Cisco PrR/OER won't help in this situation either.
|
|
|
s*****g
发帖数: 1055 | 11 This is a great idea, let me think about it.
【在 a***n 的大作中提到】
: Borrow some idea from peer to peer protocols?
:
: No, Cisco PrR/OER won't help in this situation either.
|
z**r
发帖数: 17771 | 12 write a cutomized script to healthcheck site a and b, and give the feedback
to the wan/dns load balancer.
I used to write a script on F5 box to probe some remote site availability,
and modify the bgp settings dynamically based on the feedback. F5 still
lists it as the study case for SP failover
【在 s*****g 的大作中提到】
: How?
|
s*****g
发帖数: 1055 | 13 Load balancing is not involved, just pure point 2 point TCP connection, Can
you elaborate what problem you were trying to solve? What I really trying to
do is to build a overlay network to circumvent underlying unstable Internet
. Health checking itself is not a problem, the problem is how to feed health
check data back to OSPF so tunnel metric can be dynamically advertised, do
you such hook exist on Linux/FreeBSD? |
s*****g
发帖数: 1055 | 14 Assuming we have a bunch of data centers across different geo-locations, the
connection among those data centers are over public Internet, applications
require that data centers can talk to each other with minimum latency,
applications itself will take care of data encryption so no VPN devices are
needed. Since we have no control of the underlying transport, we can have
the following situation: there are packet drops between Tokyo and Seattle
due to some transport providers link congestion, but the paths of Tokyo <-->
LA, LA<-->Seattle are good, unfortunately we have no ability to route Tokyo
-> Seattle traffic via LA. One solution I am thinking of is to build a full
-mesh GRE tunnels among the data centers and run IGP over GRE tunnels, so I
can have some level of control of the path. The big problem of doing this is
, how do I feed the underlying network performace back and dynamically
adjust the GRE metrics? the considered parameters can be latency, packet
drop rate, throughput etc.
thoughts? I am sure many other people faced or are
facing the same problem. |
c*****i
发帖数: 631 | 15 纯ip挺难的吧。你这个问题是SDN的经典应用之一，可惜现在各家的SDN都还在beta阶段
。如果是mpls te和isis倒是可以在cspf上看看。 |
s*****g
发帖数: 1055 | 16 MPLS TE won't help, first, We don't have our own transport, second, even if
we do, MPLS TE signaling is not aware of underlying physical link
utilization.
Theatrically I don't think there is any difficulty of implementing what I
want, I just don't have the expertise to write the code.
【在 c*****i 的大作中提到】
: 纯ip挺难的吧。你这个问题是SDN的经典应用之一，可惜现在各家的SDN都还在beta阶段
: 。如果是mpls te和isis倒是可以在cspf上看看。
|
z**r
发帖数: 17771 | 17 can wan/dns load balancing resolve this?
the
applications
are
->
Tokyo
full
I
【在 s*****g 的大作中提到】
: Assuming we have a bunch of data centers across different geo-locations, the
: connection among those data centers are over public Internet, applications
: require that data centers can talk to each other with minimum latency,
: applications itself will take care of data encryption so no VPN devices are
: needed. Since we have no control of the underlying transport, we can have
: the following situation: there are packet drops between Tokyo and Seattle
: due to some transport providers link congestion, but the paths of Tokyo <-->
: LA, LA<-->Seattle are good, unfortunately we have no ability to route Tokyo
: -> Seattle traffic via LA. One solution I am thinking of is to build a full
: -mesh GRE tunnels among the data centers and run IGP over GRE tunnels, so I
|
s*****g
发帖数: 1055 | 18 How?
【在 z**r 的大作中提到】
: can wan/dns load balancing resolve this?
:
: the
: applications
: are
: ->
: Tokyo
: full
: I
|
c*****i
发帖数: 631 | 19 你可以用ipsla+eem，但是我觉得不flexible和scalable。因为每个router做自己的没
有全局观，可能会出问题。SDN是最好的，但是现在还是beta，而且需要你对编程有相
当的要求。
if
【在 s*****g 的大作中提到】
: MPLS TE won't help, first, We don't have our own transport, second, even if
: we do, MPLS TE signaling is not aware of underlying physical link
: utilization.
: Theatrically I don't think there is any difficulty of implementing what I
: want, I just don't have the expertise to write the code.
|
s*****g
发帖数: 1055 | 20 Not every vendor supports ipsla monitoring, I am actually thinking of doing
this on Linux.
SDN is far from production, and the way I understand SDN, the controller has
to be close to the box so you can use OOB network for flow configuration, I
think Google is doing some kind of SDN but with off-line computation?
【在 c*****i 的大作中提到】
: 你可以用ipsla+eem，但是我觉得不flexible和scalable。因为每个router做自己的没
: 有全局观，可能会出问题。SDN是最好的，但是现在还是beta，而且需要你对编程有相
: 当的要求。
:
: if
|
|
|
a***n
发帖数: 262 | 21 So each site is multi-homed to the same or different ISPs?
In either case you won't have control
over the ISP's internal routing since you mentioned over public internet
and you might have control over the exit/entry point from/to you?
Cisco PfR if Cisco?
http://www.cisco.com/en/US/products/ps8787/products_ios_protoco
or put certain prefixes in different VRF on each site to measure the
performance of different path or PBR, then adjust your internal routing in
Linux?
doing
has
I
【在 s*****g 的大作中提到】
: Not every vendor supports ipsla monitoring, I am actually thinking of doing
: this on Linux.
: SDN is far from production, and the way I understand SDN, the controller has
: to be close to the box so you can use OOB network for flow configuration, I
: think Google is doing some kind of SDN but with off-line computation?
|
s*****g
发帖数: 1055 | 22 No, Cisco PrR/OER won't help in this situation either.
【在 a***n 的大作中提到】
: So each site is multi-homed to the same or different ISPs?
: In either case you won't have control
: over the ISP's internal routing since you mentioned over public internet
: and you might have control over the exit/entry point from/to you?
: Cisco PfR if Cisco?
: http://www.cisco.com/en/US/products/ps8787/products_ios_protoco
: or put certain prefixes in different VRF on each site to measure the
: performance of different path or PBR, then adjust your internal routing in
: Linux?
:
|
a***n
发帖数: 262 | 23 Borrow some idea from peer to peer protocols?
No, Cisco PrR/OER won't help in this situation either.
【在 s*****g 的大作中提到】
: No, Cisco PrR/OER won't help in this situation either.
|
s*****g
发帖数: 1055 | 24 This is a great idea, let me think about it.
【在 a***n 的大作中提到】
: Borrow some idea from peer to peer protocols?
:
: No, Cisco PrR/OER won't help in this situation either.
|
z**r
发帖数: 17771 | 25 write a cutomized script to healthcheck site a and b, and give the feedback
to the wan/dns load balancer.
I used to write a script on F5 box to probe some remote site availability,
and modify the bgp settings dynamically based on the feedback. F5 still
lists it as the study case for SP failover
【在 s*****g 的大作中提到】
: How?
|
s*****g
发帖数: 1055 | 26 Load balancing is not involved, just pure point 2 point TCP connection, Can
you elaborate what problem you were trying to solve? What I really trying to
do is to build a overlay network to circumvent underlying unstable Internet
. Health checking itself is not a problem, the problem is how to feed health
check data back to OSPF so tunnel metric can be dynamically advertised, do
you such hook exist on Linux/FreeBSD? |
z**r
发帖数: 17771 | 27 I was resolving physcial site failover based on the availability of certain
applications. It's similar to what you are trying to resolve, right?
I happened to have F5 box next to the Internet Routers, so I put a script on
F5 box, and the script does the healthcheck and dynamically modify the BGP
stuff.
So in your situation, you can probably use the TCL script built-in in the
router to do the healthcheck and modify the OSPF config, say metrics? BTW,
you run OSPF with your SP?
Can
to
Internet
health
do
【在 s*****g 的大作中提到】
: Load balancing is not involved, just pure point 2 point TCP connection, Can
: you elaborate what problem you were trying to solve? What I really trying to
: do is to build a overlay network to circumvent underlying unstable Internet
: . Health checking itself is not a problem, the problem is how to feed health
: check data back to OSPF so tunnel metric can be dynamically advertised, do
: you such hook exist on Linux/FreeBSD?
|
s*****g
发帖数: 1055 | 28 No, I run BGP with SPs, I plan to run OSPF on the overlay network (GRE
tunnels among data centers), lots of work needs to be done even if this idea
works, for example, how to avoid route oscillations? how much can you
trust the probes?
certain
on
BGP
【在 z**r 的大作中提到】
: I was resolving physcial site failover based on the availability of certain
: applications. It's similar to what you are trying to resolve, right?
: I happened to have F5 box next to the Internet Routers, so I put a script on
: F5 box, and the script does the healthcheck and dynamically modify the BGP
: stuff.
: So in your situation, you can probably use the TCL script built-in in the
: router to do the healthcheck and modify the OSPF config, say metrics? BTW,
: you run OSPF with your SP?
:
: Can
|
z**r
发帖数: 17771 | 29 ok. then you can probably use the same method that I used.
if the probe runs over the data path of the traffic, it would be perfectly
fine to trust it as long as it's well written. I used the similar way to BGP
signaling to probe the remote application ...
idea
【在 s*****g 的大作中提到】
: No, I run BGP with SPs, I plan to run OSPF on the overlay network (GRE
: tunnels among data centers), lots of work needs to be done even if this idea
: works, for example, how to avoid route oscillations? how much can you
: trust the probes?
:
: certain
: on
: BGP
|
s*****g
发帖数: 1055 | 30 This is what I need:
http://nms.lcs.mit.edu/ron/
BGP
【在 z**r 的大作中提到】
: ok. then you can probably use the same method that I used.
: if the probe runs over the data path of the traffic, it would be perfectly
: fine to trust it as long as it's well written. I used the similar way to BGP
: signaling to probe the remote application ...
:
: idea
|
|
|
n******t
发帖数: 4406 | 31 我没有搞懂这个为什么是SDN的经典应用，我怎么觉得这是一网络性能优化的问题。
【在 c*****i 的大作中提到】
: 纯ip挺难的吧。你这个问题是SDN的经典应用之一，可惜现在各家的SDN都还在beta阶段
: 。如果是mpls te和isis倒是可以在cspf上看看。
|
