hardware recommandation for internet scale route-reflectors
Hello, I'm looking for help on sizing hardware to use BIRD as route-reflectors. We have multiple routers with N x Transit providers (internet fullview v4/v6) and N x IXPs. We plan to replace our existing HW based on vendor routers to Software based routers. Obvious reason is performance/compute BGP updates on x86 boxes are supposed to be much better than limited router CPU/RAM. We want also to add some diversity in our RR setup, choice is not made yet but we plan to try routing VM from Cisco/Arista/Juniper vRR solution on one side and BIRD on the other side. 2 software providers means more work... but we suppose that if a bug is found on one side, the other one won't have it. Quick specs: - AF v4/v6 unicast, v4/v6 flowspec . - BGP add-path. - 50/100 peers, some are edge routers with internet scale, other are internal routers with only a limite subset of routes. - We want to implement BMP in a near future (does not seems avail in BIRD yet) I wonder if this would fit for BIRD router box, with some room for future growth in the DFZ. - CPU: Xeon gold 3,2GHz 8cores/16threads - RAM: 32 or 64gb How would BIRD behave with such HW specs ? is it overkill ? Thanks in advance for you help/comment Nico
Hello!
I wonder if this would fit for BIRD router box, with some room for future growth in the DFZ.
- CPU: Xeon gold 3,2GHz 8cores/16threads
- RAM: 32 or 64gb
How would BIRD behave with such HW specs ? is it overkill ?
The memory is probably overkill for RR in single-table setup, you'll fit in 4G with a large margin. CPU is OK. With no or minimal filters, also the convergence time isn't going to be more than several minutes on this scale. With the multithreaded version (now 3.0-alpha0), you're gonna get convergence times in tens of seconds, probably. Maria
Hello Maria, Le 08/11/2022 à 16:50, Maria Matejka a écrit :
Hello!
I wonder if this would fit for BIRD router box, with some room for future growth in the DFZ.
- CPU: Xeon gold 3,2GHz 8cores/16threads
- RAM: 32 or 64gb
How would BIRD behave with such HW specs ? is it overkill ? The memory is probably overkill for RR in single-table setup, you'll fit in 4G with a large margin.
CPU is OK. With no or minimal filters, also the convergence time isn't going to be more than several minutes on this scale.
With the multithreaded version (now 3.0-alpha0), you're gonna get convergence times in tens of seconds, probably.
Maria
Thanks for the feedback ! that's helpful. I'll try to get one machine to make my test on then. I assume also that 3.0-alpha is not suitable for production yet :) but I'll try both anyway. Thanks and best regards! Nicolas
Hello!
I assume also that 3.0-alpha is not suitable for production yet :) but I'll try both anyway.
I'd love if you tried it. There may be some weird untested behavior, as well as some known bugs. Mostly you may run into problems with running ROA checks from CLI and also MRT is completely untested in 3.0-alpha0. We're doing our best to deliver 3.0-alpha1 soon. Maria
Hello Nicolas, We are using BIRD as route reflector in addition to Juno’s VRR (need some feature like ORR) The version that we use is 2.0.7 hosted on hypervisor and the VM has: - RAM: 24G - CPU: 8vCPU (Xeon CPU) We have around 30 IPv4 sessions and 30 IPv6 sessions (per instance of Bird) The memory used is 2G The number of routes is +9M IPv6 and +1.6M IPv6 Feature used: - Multiple table - OSPF - AF v4/v6 unicast - BGP add-path Computation and convergence are quite ok.
From my experience, we never had issue on convergence since 2 years, what we are missing is add-path with the number of route and ORR.
Best Alexandre Corso
On 9 Nov 2022, at 19:16, Nicolas Piatto via Bird-users <bird-users@trubka.network.cz> wrote:
Hello Maria,
Le 08/11/2022 à 16:50, Maria Matejka a écrit :
Hello!
I wonder if this would fit for BIRD router box, with some room for future growth in the DFZ.
- CPU: Xeon gold 3,2GHz 8cores/16threads
- RAM: 32 or 64gb
How would BIRD behave with such HW specs ? is it overkill ? The memory is probably overkill for RR in single-table setup, you'll fit in 4G with a large margin.
CPU is OK. With no or minimal filters, also the convergence time isn't going to be more than several minutes on this scale.
With the multithreaded version (now 3.0-alpha0), you're gonna get convergence times in tens of seconds, probably.
Maria
Thanks for the feedback ! that's helpful.
I'll try to get one machine to make my test on then.
I assume also that 3.0-alpha is not suitable for production yet :) but I'll try both anyway.
Thanks and best regards!
Nicolas
<OpenPGP_0x0F5B8D9CB7EE61F0.asc>
Hello Alexandre, That's very helpful numbers. I'll check when I can get my hand on a test machine. Nico Le 09/11/2022 à 13:16, Alexandre Corso a écrit :
Hello Nicolas,
We are using BIRD as route reflector in addition to Juno’s VRR (need some feature like ORR)
The version that we use is 2.0.7 hosted on hypervisor and the VM has: - RAM: 24G - CPU: 8vCPU (Xeon CPU)
We have around 30 IPv4 sessions and 30 IPv6 sessions (per instance of Bird) The memory used is 2G The number of routes is +9M IPv6 and +1.6M IPv6
Feature used: - Multiple table - OSPF - AF v4/v6 unicast - BGP add-path
Computation and convergence are quite ok.
From my experience, we never had issue on convergence since 2 years, what we are missing is add-path with the number of route and ORR.
Best
Alexandre Corso
On 9 Nov 2022, at 19:16, Nicolas Piatto via Bird-users <bird-users@trubka.network.cz> wrote:
Hello Maria,
Le 08/11/2022 à 16:50, Maria Matejka a écrit :
Hello!
I wonder if this would fit for BIRD router box, with some room for future growth in the DFZ.
- CPU: Xeon gold 3,2GHz 8cores/16threads
- RAM: 32 or 64gb
How would BIRD behave with such HW specs ? is it overkill ? The memory is probably overkill for RR in single-table setup, you'll fit in 4G with a large margin.
CPU is OK. With no or minimal filters, also the convergence time isn't going to be more than several minutes on this scale.
With the multithreaded version (now 3.0-alpha0), you're gonna get convergence times in tens of seconds, probably.
Maria
Thanks for the feedback ! that's helpful.
I'll try to get one machine to make my test on then.
I assume also that 3.0-alpha is not suitable for production yet :) but I'll try both anyway.
Thanks and best regards!
Nicolas
<OpenPGP_0x0F5B8D9CB7EE61F0.asc>
participants (3)
-
Alexandre Corso -
Maria Matejka -
Nicolas Piatto