RE: MPLS in metro access networks

Mailman List Mirror (Read Only)
1

I guess you answered your own question: "And I'm not sure what faster
switching/routing has to do with MPLS:)"

As far as CEF and such goes, I couldn't disagree with that (as I was not
comparing MPLS to other optimized forwarding techniques), however, MPLS is
not a vendor-proprietary forwarding mechanism, which means that I can deploy
it worldwide, or state-wide, whatever the case may be, in my network and
have the benefit of using only ONE protocol with MPLS-enabled/aware
routers/switches. A definate plus over the other proprietary fast switching
techniques you mentioned.

Your last statement indicates "added services" have nothing to do with the
the fast switching processing of MPLS, when in fact these services depend
upon the faster delivery of the non-proprietary fast switching of MPLS. As
quoted from the rfc:

"This memo presents an approach for building core Virtual Private
   Network (VPN) services in a service provider's MPLS backbone. This
   approach uses Multiprotocol Label Switching (MPLS) running in the
   backbone to provide premium services in addition to best effort
   services."

Marc

2

Maybe I'm getting confused. The original post asked the question "what
motivates them" (RBOCs, ILECs, and CLECs) to implement MPLS. You answered
that fast switching/routing was a reason. I disagree with this because
designing and implementing MPLS just for speed benefits is a bit too
cumbersome and complex compared to using local caching mechanisms that are
just as fast, if not faster. Saying that using MPLS as an alternative to
using local caching mechanisms because of standardization doesn't make sense
to me either because the local caching mechanisms are in place regardless.
In fact, you can't run MPLS on most vendor hardware without running their
proprietary caching (Cisco mandates using CEF before implementing MPLS and
Juniper uses it's FPC hardware architecture regardless of MPLS). So to add
to my point, there is no speed benefit in running MPLS if you are already
using modern caching techniques, which most service providers interested in
MPLS are already doing.

To respond to your second point regarding using added services I agree
completely that these services require MPLS labels to work. However, this
still has nothing to do with speed benefits. You say "these services depend
upon the faster delivery" of MPLS but the RFC doesn't mention speed at all.
It just says "This approach uses MPLS running in the backbone to provide
premium services". Any MPLS added service uses label stacking which allows
for the RFC stated "premium services".

Cheers,

-Michael Cohen

3

Srihari:

I'm going to attempt to answer your original question:

-- If so, what motivates them to do so? Any analysis of the driving
forces is appreciated.

All:

I've been consulting MPLS solutions for service providers here in Asia.
The main reason for running MPLS is traffic engineering. The main
reason for desiring traffic engineering is because of the
hyper-aggregation problem with the routing protocols. I talked to one
provider who was obviously stressed and showed emotion when addressing
this problem. Basically, they appeared just shy of desperate.

You also have more flexibility separating the data and control planes.
Remember that the control plane (RSVP, IP Telephony, IBGP sessions
across the network core) follows the same best path(s) as the data plan.
Fortunately for RSVP, you can strict and loose route to avoid this
problem. I assume this is important because RSVP is how you signal the
LSP's to solve the original problem! You can engineer each plane and
classify flows into Forward Equivalency Classes (FECs). Think of the
example I will post below as the endpoints to the L2 VPN going towards
the Virtual Circuit (VC FEC).

WARNING: Traffic Engineering can create more problems than it solves if
you do not engineer correctly. Data collection BEFORE traffic
engineering is essential. You should enter step 2 (configuring TE) with
full understanding of your network.

How do I apply this in a metro access network? Before we talk about
MPLS, let's talk about a metro network that uses Super VLAN Aggregation
(not standardized, Foundry and Extreme support as well as other vendors
who I apologize for not mentioning if they do. If you work for one of
these vendors please let me know so I can tell my customers who our
competitors are :-). You can traffic engineer using 802.1s STP
groupings, but you have scalability limitations (12 bits of VLAN ID in
.1q tag) of 4096x4096. This is an excellent solution with all of the
other enhancements such as 802.1w (Rapid Spanning Tree for failover
times of less than 5 seconds and as fast as 50ms). Additionally, vendor
proprietary standards help out such as Foundry's Super Span for
segregating STP domains and tunneling BPDU's to arrive where they
should.

Now let's talk about MPLS. Most of you are smarter than I and familiar
with Cisco IOS. Foundry uses a similar flavor of the CLI (like
different UNIX versions :slight_smile: I'll drop a config on how you can engineer
4 VLANS through your network.

The users think they are on the same Ethernet wire (Ethernet Psuedo Wire
End-to-End Encapsulation). One LSP transmits and one LSP receives (full
duplex). The technology involved is simple Layer2 VPN Draft Martini.
It is only point to point and this is not the forum to discuss the ever
evolving standards that are competing; however, the Martini drafts are
mature and I believe ready to reach informational RFC status very soon.

I don't believe the market (NANOG folks reading this message for
example) wants the complexity of running BGP on L2 or L3 VPNs. This is
why you won't find too much support for the new draft Kompella (not to
be confused with the developing VPLS draft VKompella). Juniper showed a
configuration slide of this at MPLS Japan last week (original draft
recently released and not mature "00" with no revisions and missing
information). I would request someone from Juniper post a configuration
to this list showing my same example with Kompella support for
complexity comparison. If I were a Juniper customer already running L3
VPN I would take a serious look at this. It reduces the complexity a
bit, but not enough for my liking.

Everyone agrees that separating MAC address learning is essential for
solving MAC address scalability requirements. Targeted LDP sessions
appear to be a lot less complex than BGP and we may see some solutions
arrive soon. The Martini drafts got the LDP targeted sessions covered
nicely. I'm an SE who reads NANOG to help better understand my
customers' problems. I'm not a developer and I'm not caught up in the
whirlwind going on in the MPLS WG or back in San Jose. I'll leave that
task to smarter folks with better access to Starbucks.

I also encourage you to take a look at using Super VLAN technology with
an MPLS core to solve scalability problems when you want the best of
both words.

Example Terminology:

LSP: Label Switched Path that uses the tunnel label. Signaled and
setup through RSVP in this example.

VLL: Virtual Leased Line that uses the VC label.

OSPF-TE: Opaque LSA support. Note that I've configured and I HIGHLY
recommend you do this whether you use CSPF or not. "show ip ospf
database link opaque" can be an awesome command and I believe it is the
same on Cisco IOS. For those of you that have this ability and
understanding, you may want to generate these LSA's regardless of your
plans for MPLS.

Hot standby paths: Pre-signaled RSVP paths for faster failover in this
example.

My example also uses statically defined VC label definitions. This
brings up the VC as soon as the LSP route comes up (show mpls route) and
reduces convergence time by eliminating the need for targeted LDP
overhead. NOTE: Consider this L2 VPN complexity that can be eliminated
if you think it is too complex to manage. These are the values you see
after the vll-peer statements.

I will demonstrate the mixing and matching of tagged 802.1Q lines as
well.

We have a topology as follows:

   B
A_/ \_D
  \ /
   C

Please don't give me any feedback of "Why" I would run MPLS on such a
simple network. It is meant for simplicity only. Note that these are
working configs for the folks that knock MPLS for configuration
complexity. I cannot speak for other vendors. You would also need our
latest and greatest code (7.5.00) which is not available to our non beta
customers until next week. I've included configs for B, but not C. I'm
sure you can all figure out C.

4

The only problem with this approach is the idea that you can divource a
standards based approach like MPLS with the vendor specific implimentation
of switching methodologies. On the Cisco platform, for example, MPLS is
deeply tied into CEF, for somewhat logical reasons - Cisco saw no need to
reinvent the wheel, and didn't want to go down the path of a new forwarding
technology.

The platforms are STILL using proprietary switching methods - they always
will, as it is a distiguisher.

You seem to imply that if you use MPLS, you won't be using other methods of
distributed or fast switching on your routers. This simply is not the case.
On Cisco's you will still use CEF, and on Junipers you will use ASIC-based
FPC switching.

MPLS is not useful in and of itself as a switching mechanism. However, it is
useful for TE, VPNs, etc. If you enable MPLS on your network to get "better
performance", "faster speeds", or a "more reliable core", you will be
disappointed in the end, as the performance is the same, speed is the same,
and reliability is lower due to immature code.

- Daniel Golding

5

Forgive me for the config error on the previous post. Loopbacks on
router D should have been X.X.50.X not X.X.5.X.

Gary

From: owner-nanog@merit.edu [mailto:owner-nanog@merit.edu] On Behalf

Of

Gary Blankenship
Sent: Friday, November 16, 2001 5:34 PM
To: nanog@merit.edu
Subject: RE: MPLS in metro access networks

Srihari:

I'm going to attempt to answer your original question:

> -- If so, what motivates them to do so? Any analysis of the driving
> forces is appreciated.

All:

I've been consulting MPLS solutions for service providers here in

Asia.

The main reason for running MPLS is traffic engineering. The main
reason for desiring traffic engineering is because of the
hyper-aggregation problem with the routing protocols. I talked to one
provider who was obviously stressed and showed emotion when addressing
this problem. Basically, they appeared just shy of desperate.

You also have more flexibility separating the data and control planes.
Remember that the control plane (RSVP, IP Telephony, IBGP sessions
across the network core) follows the same best path(s) as the data

plan.

Fortunately for RSVP, you can strict and loose route to avoid this
problem. I assume this is important because RSVP is how you signal

the

LSP's to solve the original problem! You can engineer each plane and
classify flows into Forward Equivalency Classes (FECs). Think of the
example I will post below as the endpoints to the L2 VPN going towards
the Virtual Circuit (VC FEC).

WARNING: Traffic Engineering can create more problems than it solves

if

you do not engineer correctly. Data collection BEFORE traffic
engineering is essential. You should enter step 2 (configuring TE)

with

full understanding of your network.

How do I apply this in a metro access network? Before we talk about
MPLS, let's talk about a metro network that uses Super VLAN

Aggregation

(not standardized, Foundry and Extreme support as well as other

vendors

who I apologize for not mentioning if they do. If you work for one of
these vendors please let me know so I can tell my customers who our
competitors are :-). You can traffic engineer using 802.1s STP
groupings, but you have scalability limitations (12 bits of VLAN ID in
.1q tag) of 4096x4096. This is an excellent solution with all of the
other enhancements such as 802.1w (Rapid Spanning Tree for failover
times of less than 5 seconds and as fast as 50ms). Additionally,

vendor

proprietary standards help out such as Foundry's Super Span for
segregating STP domains and tunneling BPDU's to arrive where they
should.

Now let's talk about MPLS. Most of you are smarter than I and

familiar

with Cisco IOS. Foundry uses a similar flavor of the CLI (like
different UNIX versions :slight_smile: I'll drop a config on how you can

engineer

4 VLANS through your network.

The users think they are on the same Ethernet wire (Ethernet Psuedo

Wire

End-to-End Encapsulation). One LSP transmits and one LSP receives

(full

duplex). The technology involved is simple Layer2 VPN Draft Martini.
It is only point to point and this is not the forum to discuss the

ever

evolving standards that are competing; however, the Martini drafts are
mature and I believe ready to reach informational RFC status very

soon.

I don't believe the market (NANOG folks reading this message for
example) wants the complexity of running BGP on L2 or L3 VPNs. This

is

why you won't find too much support for the new draft Kompella (not to
be confused with the developing VPLS draft VKompella). Juniper showed

a

configuration slide of this at MPLS Japan last week (original draft
recently released and not mature "00" with no revisions and missing
information). I would request someone from Juniper post a

configuration

to this list showing my same example with Kompella support for
complexity comparison. If I were a Juniper customer already running

L3

VPN I would take a serious look at this. It reduces the complexity a
bit, but not enough for my liking.

Everyone agrees that separating MAC address learning is essential for
solving MAC address scalability requirements. Targeted LDP sessions
appear to be a lot less complex than BGP and we may see some solutions
arrive soon. The Martini drafts got the LDP targeted sessions covered
nicely. I'm an SE who reads NANOG to help better understand my
customers' problems. I'm not a developer and I'm not caught up in the
whirlwind going on in the MPLS WG or back in San Jose. I'll leave

that

task to smarter folks with better access to Starbucks.

I also encourage you to take a look at using Super VLAN technology

with

an MPLS core to solve scalability problems when you want the best of
both words.

Example Terminology:

LSP: Label Switched Path that uses the tunnel label. Signaled and
setup through RSVP in this example.

VLL: Virtual Leased Line that uses the VC label.

OSPF-TE: Opaque LSA support. Note that I've configured and I HIGHLY
recommend you do this whether you use CSPF or not. "show ip ospf
database link opaque" can be an awesome command and I believe it is

the

same on Cisco IOS. For those of you that have this ability and
understanding, you may want to generate these LSA's regardless of your
plans for MPLS.

Hot standby paths: Pre-signaled RSVP paths for faster failover in

this

example.

My example also uses statically defined VC label definitions. This
brings up the VC as soon as the LSP route comes up (show mpls route)

and

reduces convergence time by eliminating the need for targeted LDP
overhead. NOTE: Consider this L2 VPN complexity that can be

eliminated

if you think it is too complex to manage. These are the values you

see

after the vll-peer statements.

I will demonstrate the mixing and matching of tagged 802.1Q lines as
well.

We have a topology as follows:

   B
A_/ \_D
  \ /
   C

Please don't give me any feedback of "Why" I would run MPLS on such a
simple network. It is meant for simplicity only. Note that these are
working configs for the folks that knock MPLS for configuration
complexity. I cannot speak for other vendors. You would also need

our

latest and greatest code (7.5.00) which is not available to our non

beta

customers until next week. I've included configs for B, but not C.

I'm

sure you can all figure out C.

--------------------------------------------------------
hostname A

router mpls
policy
  traffic-eng ospf

mpls-interface p2/1
mpls-interface p2/2

path B-to-D
  strict 192.168.10.2
  strict 192.168.20.1
path C-to-D
  strict 192.168.30.2
  strict 192.168.40.1

lsp LSP1
  to 192.168.50.1
  primary B-to-D
  secondary C-to-D
    standby
  enable

lsp LSP2
  to 192.168.50.2
  primary B-to-D
  secondary C-to-D
    standby
  enable

lsp LSP3
  to 192.168.50.3
  primary B-to-D
  secondary C-to-D
    standby
  enable

lsp LSP4
  to 192.168.50.4
  primary C-to-D
  secondary B-to-D
    standby
  enable

vll VLAN10 10
  vll-peer 192.168.50.1 800001 900001
  vlan 10
   tagged e 1/1
vll VLAN20 20
  vll-peer 192.168.50.2 800002 900002
  vlan 20
   tagged e 1/1
vll VLAN30 30
  vll-peer 192.168.50.3 800003 900003
  vlan 30
   tagged e 1/1
vll VLAN40 40
  vll-peer 192.168.50.4 800004 900004
  vlan 40
   tagged e 1/1
!
boot sys slot1 N2M07500.bin
route-only
router ospf
area 0
!
interface loopback 1
ip address 192.168.5.1 255.255.255.255
ip ospf area 0

interface loopback 2
ip address 192.168.5.2 255.255.255.255
ip ospf area 0

interface loopback 3
ip address 192.168.5.3 255.255.255.255
ip ospf area 0

interface loopback 4
ip address 192.168.5.4 255.255.255.255
ip ospf area 0
!
interface ethernet 1/1
enable
!
interface pos 2/1
enable
ip address 192.168.10.1 255.255.255.0
ip ospf area 0
!
interface pos 2/2
enable
ip address 192.168.30.1 255.255.255.0
ip ospf area 0
!
!
!
!
!
end

--------------------------------------------

hostname D
router mpls
policy
  traffic-eng ospf

mpls-interface p2/1
mpls-interface p2/2

path B-to-A
  strict 192.168.20.2
  strict 192.168.10.1
path C-to-A
  strict 192.168.40.2
  strict 192.168.30.1

lsp LSP1
  to 192.168.5.1
  primary B-to-A
  secondary C-to-A
    standby
  enable

lsp LSP2
  to 192.168.5.2
  primary B-to-A
  secondary C-to-A
    standby
  enable

lsp LSP3
  to 192.168.5.3
  primary B-to-A
  secondary C-to-A
    standby
  enable

lsp LSP4
  to 192.168.5.4
  primary C-to-A
  secondary B-to-A
    standby
  enable

vll VLAN10 10
  vll-peer 192.168.5.1 900001 800001
   untagged e 1/1
vll VLAN20 20
  vll-peer 192.168.5.2 900002 800002
   untagged e 1/2
vll VLAN30 30
  vll-peer 192.168.5.3 900003 800003
   untagged e 1/3
vll VLAN40 40
  vll-peer 192.168.5.4 900004 800004
   untagged e 1/4

!
boot sys slot1 N2M07500.bin
route-only
router ospf
area 0
!
interface loopback 1
ip address 192.168.5.1 255.255.255.255
ip ospf area 0

interface loopback 2
ip address 192.168.5.2 255.255.255.255
ip ospf area 0

interface loopback 3
ip address 192.168.5.3 255.255.255.255
ip ospf area 0

interface loopback 4
ip address 192.168.5.4 255.255.255.255
ip ospf area 0
!
interface ethernet 1/1
enable
!
interface pos 2/1
port-name NI800_D_POS_3/1
enable
ip address 192.168.40.1 255.255.255.0
ip ospf area 0
!
interface pos 2/2
port-name NI800_D_POS_3/2
enable
ip address 192.168.20.1 255.255.255.0
ip ospf area 0!
!
!
!
!
end

-------------------------------------------------
hostname C

router mpls
policy
  admin-group OSPF-Primary-Link 2
  traffic-eng ospf

mpls-interface p6/1
  admin-group 2
mpls-interface p6/2
  admin-group 2

!
boot sys slot1 N2M07500B5.bin
route-only
router ospf
area 0
!
interface pos 6/1
enable
ip address 192.168.40.2 255.255.255.0
ip ospf area 0
!
interface pos 6/2
enable
ip address 192.168.30.2 255.255.255.0
ip ospf area 0
!
!
!
!
!
end

Gary Blankenship
Foundry Networks
Systems Engineer - Japan
CCIE #5009
garyb@foundrynet.com
www.foundrynet.com

PC Magazine Selects Foundry FastIron 4802 as Best High End Ethernet
Switch:
http://www.pcmag.com/article/0,2997,s%3D25109%26a%3D17629,00.asp

> From: owner-nanog@merit.edu [mailto:owner-nanog@merit.edu] On Behalf
Of
> Michael Cohen
> Sent: Friday, November 16, 2001 3:45 AM
> To: nanog@merit.edu
> Subject: RE: MPLS in metro access networks
>
>
> Maybe I'm getting confused. The original post asked the question
"what
> motivates them" (RBOCs, ILECs, and CLECs) to implement MPLS. You
answered
> that fast switching/routing was a reason. I disagree with this
because
> designing and implementing MPLS just for speed benefits is a bit too
> cumbersome and complex compared to using local caching mechanisms

that

are
> just as fast, if not faster. Saying that using MPLS as an

alternative

to
> using local caching mechanisms because of standardization doesn't

make

> sense
> to me either because the local caching mechanisms are in place
regardless.
> In fact, you can't run MPLS on most vendor hardware without running
their
> proprietary caching (Cisco mandates using CEF before implementing

MPLS

and
> Juniper uses it's FPC hardware architecture regardless of MPLS). So
to
> add
> to my point, there is no speed benefit in running MPLS if you are
already
> using modern caching techniques, which most service providers
interested
> in
> MPLS are already doing.
>
> To respond to your second point regarding using added services I

agree

> completely that these services require MPLS labels to work.

However,

this
> still has nothing to do with speed benefits. You say "these

services

> depend
> upon the faster delivery" of MPLS but the RFC doesn't mention speed

at

all.
> It just says "This approach uses MPLS running in the backbone to
provide
> premium services". Any MPLS added service uses label stacking which
> allows
> for the RFC stated "premium services".
>
> Cheers,
>
> -Michael Cohen
>
> From: Quibell, Marc [mailto:mquibell@icn.state.ia.us]
> Sent: Thursday, November 15, 2001 12:04 PM
> To: 'mcohen@thrupoint.net'; nanog@merit.edu
> Subject: RE: MPLS in metro access networks
>
>
> I guess you answered your own question: "And I'm not sure what

faster

> switching/routing has to do with MPLS:)"
>
> As far as CEF and such goes, I couldn't disagree with that (as I was
not
> comparing MPLS to other optimized forwarding techniques), however,
MPLS is
> not a vendor-proprietary forwarding mechanism, which means that I

can

> deploy
> it worldwide, or state-wide, whatever the case may be, in my network
and
> have the benefit of using only ONE protocol with MPLS-enabled/aware
> routers/switches. A definate plus over the other proprietary fast
> switching
> techniques you mentioned.
>
> Your last statement indicates "added services" have nothing to do

with

the
> the fast switching processing of MPLS, when in fact these services
depend
> upon the faster delivery of the non-proprietary fast switching of
MPLS. As
> quoted from the rfc:
>
> "This memo presents an approach for building core Virtual Private
> Network (VPN) services in a service provider's MPLS backbone.

This

> approach uses Multiprotocol Label Switching (MPLS) running in the
> backbone to provide premium services in addition to best effort
> services."
>
> Marc
>
>
> From: Michael Cohen [mailto:mcohen@thrupoint.net]
> Sent: Thursday, November 15, 2001 11:20 AM
> To: nanog@merit.edu
> Subject: RE: MPLS in metro access networks
>
>
>
> I still have to disagree that MPLS results in faster

switching/routing

in
> modern service provider networks. Modern vendor caching mechanisms
are
> just
> as fast if not faster than MPLS processing. With the small overhead
of
> MPLS
> labels and LDP I highly doubt that you're getting any performance
increase
> over Cisco's CEF or Juniper's FPC architecture. I also doubt that
speed
> is
> a benefit that service providers consider when deciding whether or

not

> they
> want to implement MPLS. Added services that run on top of MPLS like
VPNs,
> traffic engineering, and fast rerouting capabilities (all mentioned

in

the
> original post) are more likely the benefits considered. Perhaps

when

> label
> switching was first being marketed (Ipsilon and Cisco in 1996) there
were
> some speed benefits but now I think it's the services that use MPLS
that
> are
> the major benefit.
>
> -Michael Cohen
>
> From: Quibell, Marc [mailto:mquibell@icn.state.ia.us]
> Sent: Thursday, November 15, 2001 10:59 AM
> To: 'mcohen@thrupoint.net'; 'nanog@merit.edu'
> Subject: RE: MPLS in metro access networks
>
>
> soooo...Label switching assigns labels to packet headers which

results

in
> less time and processing looking up routes, and instead relies upon

a

> label
> index for forwarding decisions? Hence my statement "faster
> switching/routing
> and less processing":slight_smile:
>
> Marc
>
>
>
> From: Michael Cohen [mailto:mcohen@thrupoint.net]
> Sent: Thursday, November 15, 2001 10:56 AM
> To: Quibell, Marc
> Subject: RE: MPLS in metro access networks
>
>
> I hope so:)
>
> From: Quibell, Marc [mailto:mquibell@icn.state.ia.us]
> Sent: Thursday, November 15, 2001 10:46 AM
> To: 'mcohen@thrupoint.net'; nanog@merit.edu
> Subject: RE: MPLS in metro access networks
>
>
> Are we talking about Multiprotocol Label Switching?
>
> Marc
>
>
> From: Michael Cohen [mailto:mcohen@thrupoint.net]
> Sent: Thursday, November 15, 2001 10:45 AM
> To: nanog@merit.edu
> Subject: RE: MPLS in metro access networks
>
>
>
> And I'm not sure what faster switching/routing has to do with MPLS:)
I
> believe one of the ideas behind MPLS benefiting metro access

networks

is
> using MPLS to deliver layer 2 VPNs across an MPLS enabled core thus
> simulating leased lines for access clients...but I'm sure somebody
will
> correct me if I'm wrong. There seems to be some hype for Martini
draft
> VPNs
> and large enterprise customers in metro areas.
>
> Cheers,
>
> -Michael Cohen
>
> From: owner-nanog@merit.edu [mailto:owner-nanog@merit.edu]On Behalf

Of

> Quibell, Marc
> Sent: Thursday, November 15, 2001 10:20 AM
> To: 'srihari varada'; nanog@merit.edu
> Subject: RE: MPLS in metro access networks
>
>
>
> I would think faster switching/routing and less processing would be
wanted
> in any mid-to-large sized network...I'm not sure what load balancing
and
> fault restoration has to do with MPLS....
>
> Marc
>
>
>
> From: srihari varada [mailto:varada@txc.com]
> Sent: Thursday, November 15, 2001 10:12 AM
> To: nanog@merit.edu
> Subject: MPLS in metro access networks
>
>
> Hello:
>
> I have heard some stressing the role of MPLS in metro access

networks.

> It is difficult for me to visualize the need for it in them while it
> is not so difficult to understand the utility (load balancing and
fault
> restoration etc.) of it in the metro backbone networks.
>
> To characterize metro access networks in the context, the following

is

> provided:
> -- aggregates traffic from residential (arriving via broadband

access

> links such as xDSL, Cable) and business consumers (arriving via
> broadband access links such as
> xDSL and high speed links such as Ethernet or SONET)
> -- funnels aggregated traffic to metro backbone networks for
destination
>
> hosts in the local metro region or remote regions across the
> internet regional
> and backbone networks. Majority of such access networks are
SONET/ATM
> based (I didn't come
> across any case of Gig Ethernet. However, I do not preculde it).
>
> Thus, there are two questions:
> -- Are there known RBOCs/ILECs and CLECs entrenching MPLS in the

said

6

MPLS is not useful in and of itself as a switching mechanism. However, it is
useful for TE, VPNs, etc. If you enable MPLS on your network to get "better
performance", "faster speeds", or a "more reliable core", you will be
disappointed in the end, as the performance is the same, speed is the same,
and reliability is lower due to immature code.

How old is your information?

The company I'm working for has been using MPLS for some 2 years now
(maybe more, maybe less, I forget). We have found the code to be quite
stable, although that speaks as much to our vendors improved QA as it does
to our internal QA and testing that takes place months before deployment.

MPLS-TE is definitely a valuable feature. It was worth it for us to deploy
it back in '99, and it's worth it now. MPLS-based VPN's are interesting
in their own right.

I would not consider MPLS to be bleeding edge anymore...and for us, it
doesn't really feel cutting edge anymore either. It's the newer features
still in the standards process that might be dangerous to deploy at this
stage, but that's what labs are for (that would be an actual lab, not
the "production lab" :-).

That being said, I would love to see some quantifiable data showing the
differences in switching latencies between labels and packets. %age increase
in efficient use of network capacity will of course depend on your backbone
design. Reliability, in the end, has more to do with organization processes
for engineering and operating your network than the quality of a vendor's
latest GA release of code.

Dave

7

If there is a difference in latency between MPLS and IP switching
you need to beat your vendor, and/or change vendors.

8

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Dave,

Are you suggesting that MPLS is truly a faster switching paradigm
across a large network? There is basically no meaurable latency
difference in MPLS networks, unless you have avoided congestion by
using TE.

As far as reliability - your milage may vary. MPLS-TE, which you
mention, is certainly one of the various vendors' more stable MPLS
features. However, just in the past 6 months, major vendors have
shipped several versions of code, (including in the "stable" code
trains), with major MPLS issues - many of which are of the "drop
packets on the floor, because we don't know what to do with them"
variety.

I don't think anyone is slamming MPLS or it's implementations
just because their might be a bug in "a vendor's latest GA release
of code". MPLS is a complexity added onto already complex code.

However, the proponents of MPLS based networks stress things like
it's reliability, speed and simplicity, over existing, pure IP
network designs. My point, is that it is certain not faster, not any
simpler, and, for many, has been considerably less reliable. I
reiterate that the point of implimenting MPLS is to be able to
provide the services it enables (of which, TE is a great example,
along with VPNs).

Of course, different levels of perceived reliability may be due to
different depths of MPLS implimentation. A design consistint of fully
meshed, explicit TE tunnels would probably be pretty reliable, by
this point. A network with many L2VPNs and RFC2547 VPNs running over
LSPs instantiated by LDP, might be less so, depending on the
platform.

- - Daniel Golding

9

However, the proponents of MPLS based networks stress things like
it's reliability, speed and simplicity, over existing, pure IP
network designs.

We do? Not all of us...:slight_smile:

Reliability? I've never heard claims that MPLS is any more or less
reliable than IP. Certainly there's fast reroute mechanisms so that a
failure can be worked around pretty quickly, but we're talking about a
failure at a lower level (link failure) or a box/LC failure (crash).
The MPLS software implementation from a given vendor should be as
reliable as the IP software; there's nothing fundamentally different
in MPLS that makes it easier to code, or inherently bug-free or any of
that stuff.

Speed?
Back in the early days, it was thought that a 20-bit label lookup
would be faster than a 32-bit IP DA lookup. With the advent of fast
hardware switching, this turns out not to be true. Speaking as I can
for only one vendor, I can tell you that anyone from my company who
claims MPLS is substantially faster at a switching lookup than IP
needs to be beat with a clue stick.

It turns out, on some platforms, that imposing a label stack is a wee
bit slower than switching an IP packet and that swapping a label is a
wee bit faster than switching an IP packet, but "wee bit" means
something in the area of a few percent. And this does not apply to
all platforms.

Simplicitly? Possibly. It might be argued that no longer needing BGP
in the core makes your network simpler. But IMHO a BGP-free core in
and of itself is not enough of a reason to deploy MPLS.

The things I think MPLS buys you are:

- traffic engineering for your core - being able to forward traffic
down a path other than the one your IGP would select, in a more
granular method and less error-prone fashion than other tools to do
such (static routes, GRE tunnels, juggling link costs, etc)

- ability to carry edge services (L3VPN, L2VPN) across your network.
And there's some QoS benefits in here, too, since you have a
hierarchical set of labels.

MPLS is not the *only* way to do these things, but what it can do has
proven useful to lots of people.

My point, is that it is certain not faster, not any
simpler, and, for many, has been considerably less reliable. I
reiterate that the point of implimenting MPLS is to be able to
provide the services it enables (of which, TE is a great example,
along with VPNs).

Absolutely. Let me also add that reliability tends to correlate with
maturity and deployment, so you'll only see existing implementations
get better over time.

Of course, different levels of perceived reliability may be due to
different depths of MPLS implimentation. A design consistint of fully
meshed, explicit TE tunnels would probably be pretty reliable, by
this point. A network with many L2VPNs and RFC2547 VPNs running over
LSPs instantiated by LDP, might be less so, depending on the
platform.

I think I agree with this. It's not that LDP is inherently any less
reliable than TE, but that with FRR one can protect against
lower-level failures. I think I prefer "availability" over
"reliability" in this context, since that takes into account TE's
ability to avoid some congestion that LDP (following the IGP path)
can't.

But that's just me.

eric

10

On Wed, Nov 28, 2001 at 08:17:30PM -0500, Leo Bicknell reportedly typed:

> That being said, I would love to see some quantifiable data showing the
> differences in switching latencies between labels and packets. %age increase

If there is a difference in latency between MPLS and IP switching
you need to beat your vendor, and/or change vendors.

If there is a difference, I would expect it to be in the order of low,
single-digit microseconds.