网络交换机上的 OpenBMC 和 FBOSS.pdf

1、Dean Kalla,Arista NetworksMike Choi,MetaOpenBMC and FBOSS on Network Switch System On a Chip Manages system hardware Provides out of band management and reporting REST APIs,etc Very popular in server world Supports OpenBMC-yocto open embedded linuxWhat is a BMC?Arista has been collaborating with Met

2、a on switches solutions since 2017Integrated BMC in switches that are traditionally BMC-lessMulti-NOS support:FBOSS,EOS,SONiC,etcArista/Meta Joint Development7368X4128x100GTH37388X5128x200GTH57700R4C-38P18x800G NIF,20 x800G FabricJ37700R4-128P128x800G FabricR47060X6-64PE64x800GTH57060DX4S32x400GTH3

3、The Metas hardware architecture changed from BMC-Classic to BMC-Lite in 2024.BMC-Classic(2023)BMC controls most of the HW components and features CPUs role as ASIC controller.CPU making requests to BMC for HW management.BMC-Lite(2024 Present)BMC only manages CPU(like Server appliances),and Provides

4、backup recovery path,in case CPU is down.BMC-Classic vs.BMC-Lite(1/3)-IntroBMC-Classic vs.BMC-Lite(2/3)-SW AspectArchitectureBMC-Classic(2023)BMC-Lite(2024 Present)BMC Role“System”Management“CPU”Management+System RecoverySW StructureA lot of boilerplate code,used by BASH scripts and Python3 scriptsC

5、onfig Driven Platform Services in C+.New projects require the development of JSON Config files only.ProsFailure Domain Isolation(Platform Management is separated from ASIC management.)Platform Service Upgrade is easy,but monolithic.Easy to integrate platform services into Metas infra serviceIntegrat

6、es with complicated rack systems(for AI)better.Fine-grained SW package managementConsBMC CPU is often overloaded(reaching 100%CPU usage in Meta Data Center)Not compatible with Metas Thrift RPC library Kernel Panic kills both SDK and Platform Services(But BMC can still reboot CPU)