Emulating federated machine learning using QEMU and Zephyr RTOS.
First, follow the Zephyr OS setup guide to install west and the Zephyr SDK.
From within your copy of the Zephyr source code repository, clone zfl as a repository application. This ensures that Zephyr kernel dependencies are accessible for the client build.
Run make to bootstrap the main zfl tool.
zfl uses qemu-system-i386 by default and searches for an executable with that name on your path.
The standard configuration for the QEMU board provided by zephyr can be tweaked to allocate more SRAM. This is especially needed for machine learning.
- Locate the qemu_x86 board file somewhere in
<zephyr_root>/boards/x86/qemu_x86/qemu_x86.dts - In the standard
zflclient15MB RAM is used. Configure this needed memory with:#define DT_DRAM_SIZE DT_SIZE_K(15360) - Additional (or less) memory can also be configured for
zflclientby changing these parameters inzflclient/prj.conf:
CONFIG_SRAM_SIZE=15360
CONFIG_MAIN_STACK_SIZE=8192
CONFIG_KERNEL_VM_SIZE=0x7000000
CONFIG_HEAP_MEM_POOL_SIZE=10485760
With zfl built, run the following commands from the zfl root directory to start clients running in QEMU x86.
// Build target
make client
// Start n clients
sudo ./zfl client <num_clients> <epochs> <batch_size>
With zfl built, run the following commands from the zfl root directory to start the central server.
make server
sudo ./zfl server <num_rounds> <clients_per_round>
Federated learning allows a fleet of devices to collaborate towards a globally trained machine learning model. Research has continued to produce novel federated learning algorithms to tackle different issues in FL such as heteorogenity and learning over data from non-identical distributions. Performance of these algorithms depend in part on the system parameters used in FL such as number of clients and number of passes. Furthermore, a realistic benchmark would require one to procure a large fleet of devices. This work seeks to introduce a software emulation framework to streamline the process of building a fleet of clients and configuring system parameters.
The server code comes baked in with nn.h by Tsoding.