Hi OCP Networking group,

Please find the spec and design package for the Facebook submission  Backpack- 
128x100G at the OCP site below:
http://www.opencompute.org/wiki/Networking/SpecsAndDesigns#Facebook_Backpack_-_128x100G


The spec version is v.92
Design package v0.1


The design files submitted here reflect the design that’s currently in 
production at Facebook datacenters.

Also below are answers to a couple of questions that were raised during the OCP 
community review:


-          How is OpenBMC different for a chassis based switch vs the previous 
release for Wedge100

o   There is the concept of a Master BMC in Backpack that acts as the central 
point of monitoring/reporting for power/temperature/fans for the entire chassis.

o   This is implemented on the CMM management module in Backpack.

o   The CMM implementation also uses a new BMC SOC, and anew kernel/utilities.

o   There are slave BMCs that exist on every line card as well as fabric cards. 
These are similar to the openBMC that exists on the Wedge100 switch.


-          How does the Backpack switch get provisioned/boot up?

o   Wedge100 is a single unit to provision, while for Backpack we have 25 units 
that provision independently

o    (12 SCMs + 12 LC/FC + 1 CMM). For speeding up initial provisioning all 
these cards go roughly in parallel (see ii)

o   ii) For each pair of LC/FC, SCM – LC/FC provisions first and then triggers 
a SCM provisioning job. In wedge100, this

o   is self contained in a single provisioning job.

o   iii) Flip side of provisioning is unprovisioning. This happens roughly 
independently for all 25 cards in Backpack. Exception here is

o   we unprovision SCM first followed by its paired LC/FC. We will likely break 
this serialization in future.

o   iv) Same infrastructure is used for provisioning Backpack as for wedge100. 
A simple way of seeing Backpack provisioning

o   is to see each card as a system to be provisioned unto itself.

o   v) Since the chassis is not ready to take traffic until all cards are 
provisioned, each card in FSW comes up as drained – I.e.

o   it does not announce out any traffic bearing prefix. Only once all cards 
have been provisioned and chassis has been audited

o   do we undrain it. We follow the same process for vendor FSWs even though 
there its a single SUP to provision


Please let us know if you have any questions about the package or if you are 
interested in integrating with Backpack from a software perspective or 
potentially obtaining a Backpack when they are generally available. We plan to 
bring Backpack for OCP Incubation committee review soon.



Ramesh Padmanabhan                                                              
Luis MartinGarcia
Technical Program Manager, Facebook                                Network 
Partner Engineering, Facebook


_______________________________________________
opencompute-networking mailing list
Unsubscribe: http://lists.opencompute.org/mailman/options/opencompute-networking

[email protected]
http://lists.opencompute.org/mailman/listinfo/opencompute-networking

Reply via email to