The MPLS WG Archive

Cell Relay Retreat>MPLS WG Archive>month:2002-Apr> msg00214



[Date Prev][Date Next][Thread Prev][Thread Next]  
  [Date Index][Thread Index][Author Index][Subject Index]

Memory at PE

  • From: "Miguel Angel Chana" <machana@cisco.com>
  • Date: Wed, 24 Apr 2002 16:49:31 +0200
  • Cc: <mpls@UU.NET>
  • Importance: Normal

Title: RE: Memory at PE
Hello, Jim:
 
I cannot be more in agreement with you. This kind of messages in this list are not
very useful, and I think we should not discuss particular implementations. For me
this should not be a place for manufacturers to discuss who does it better.
 
Having said that, I can't resist to comment that every architecture has its own
share of issues.
 
In a virtual router architecture, if there are N instances of routing, every
one receives more or less 1/N CPU process time and 1/N memory. Both resources
are finite, unfortunately.
 
Regards
 
 
-----Original Message-----
From: owner-mpls@UU.NET [mailto:owner-mpls@UU.NET]On Behalf Of Jim Guichard
Sent: Wednesday, April 24, 2002 4:20 PM
To: Andrew Walker; Sachin Kalra; Liu, Chia J (Charlie), ALCNS
Cc: mpls@UU.NET
Subject: RE: Memory at PE

Andrew,
 
I do not normally respond to these marketing type emails but I think it is important that we try and stay within the realms of reality. There are a whole bunch of things that contribute to scalability and to suggest that central CPU/memory is the bottleneck, or only consideration for scaling, is completely misleading. Jim
-----Original Message-----
From: owner-mpls@UU.NET [mailto:owner-mpls@UU.NET]On Behalf Of Andrew Walker
Sent: Wednesday, April 24, 2002 9:18 AM
To: Sachin Kalra; Liu, Chia J (Charlie), ALCNS
Cc: mpls@UU.NET
Subject: RE: Memory at PE

Sachin,

You have asked a very important question as it gets to the
heart of the scalability of a PE deployment. In a traditional router
architecture (e.g. Cisco, Juniper, Unisphere), although FIBs are
distributed out to individual line cards, the RIB is kept on a central
module. There is only one RIB and therefore all updates whether for the
Local RIB, or for each VRF must go through this RIB.
Obviously the RIB is divided into pieces so as to keep the
VRFs separate, but the whole thing is bounded by the
amount of memory and CPU cycles available to the routing plane.
Prior to BGP/MPLS VPNs, update performance of the RIB was
not a large concern, so the fact that it was still centralized did not matter.

However BGP/MPLS VPNs place a new burden on traditional routers.
VRFs are by definition specific to a particular subscriber network and
must only contain routes specific to that network; traditional routers
partition their RIBs and FIBs in order to maintain this separation. This
can become a problem for traditional routers since the RIB is maintained
in a central module and serviced by a single processor. These centralized
resources (memory and CPU) become a bottleneck as the number
of subscribers increases, especially when the PE-CE connections use chatty
routing protocols such as RIP or computationally intensive routing protocols
such as OSPF. For example, when using OSPF as the
CE-PE routing protocol, a Cisco router can only sustain approx 20 to 30
instances per device. With a distributed architecture, like the
CoSine IPSX platform, you can have more than 1500 such instances
per blade.

In a truly distributed device architecture, the RIBs are distributed
with the VRFs, and new resources (CPUs and Memory) can be
added as needed.  Thus you avoid the limitations imposed by
traditional routing architectures.

The base architecture of a PE device is absolutely critical in
determining how many BGP/MPLS VPNs the device can handle.
Trying to retrofit an existing architecture for a new network design
can lead to limitations that purposely designed hardware doesn't suffer from.
--------



At 12:49 PM 4/23/2002 -0400, Liu, Chia J (Charlie), ALCNS wrote:
>Sachin,
>
>There are two aspects:
>*       Memory Usage in GRP/RSP:   In the lab, we saw data of ~935 bytes per VPN prefix in VRF, compared to 500-600 bytes per internet prefix in global routing table.    I heard there is additional 60KB-70KB overhead per VRF.

>*       Memory Usage in Line Card:   We are particularly concerned about the VRF memory overhead in PSA/TLU of E2 16xOC-3 in GSR.    It looks like the number is different in different IOS releases.    I am interested in knowing if anyone has number on this.   Thanx.

>
>C.J. (Charlie) Liu
>
>
>> -----Original Message-----
>> From: Sachin Kalra [SMTP:skalra@opnet.com]
>> Sent: Tuesday, April 23, 2002 11:51 AM
>> To:   mpls@UU.NET
>> Subject:      Memory at PE
>>
>> Dear Group:
>>
>> I would appreciate if I can get answer to the following question regarding
>> BGP/MPLS VPNs (RFC2547bis)
>>
>> I understand that PE router maintains a separate VRF table for each VPN
>> site connected to it. I wanted to know if a PE also maintain separate RIBs
>> for each VPN site, apart from its main Local RIB? Or, does it maintain only
>> one single RIB?
>>
>> Actually, I was looking from the perspective of amount of memory required
>> at PE, if it has to maintain many VRFs and many RIBs.
>>
>> Thanks for your response.
>> Sachin Kalra
###################################################################################################### This email communication may contain CONFIDENTIAL INFORMATION and is intended only for the use of the intended recipients identified above.  If you are not the intended recipient of this communication, you must not use, disclose, distribute, copy or print this email. If you have received this communication in error, please immediately notify the sender by reply email, delete the communication and destroy all copies. ######################################################################################################


  • References: