How dimitrijejankov/bbts works
BarbaTOS is a distributed tensor operating system. Its principal workflow starts with a client submitting tensor commands, which are compiled into node-specific work, coordinated across the cluster, and executed using registered tensor operations and storage. The tree also shows matrix-multiplication, feed-forward-network, and XLA-kernel applications. Relationships below are limited to those visible in sampled source, SOURCE INDEX references, or the README; some application internals and runtime wiring were not sampled.
Main entry points: udf_manager.h Operation Registry · command_runner.h Command Runner · two_layer_compiler.h Two-Layer Compiler · generate_ffnn.cc Feed-Forward Workloads
Components 23
Client Applications
- Node CLI
applications/node/node_cli.cc - Matrix Workloads
applications/bmm/generate_bmm.cc - Feed-Forward Workloads
applications/ffnn/generate_ffnn.cc - XLA Kernel App
applications/xla_kernel/main.cc
Command Planning
- Parsed Commands
main/commands/parsed_command.h - Two-Layer Compiler
main/commands/two_layer_compiler.h - Execution Commands
main/commands/command.h - Command Runner
main/commands/command_runner.h
Cluster Execution
- Reservation Station
main/commands/reservation_station.h - Command Handlers
main/commands/command_handlers/command_handler_apply.cc - Coordinator
main/server/coordinator.cc - Compute Node
main/server/node.h - MPI Communication
main/communication/communicator.h
Tensor Runtime
- Tensor Model
main/tensor/tensor.h - Tensor Factory
main/tensor/tensor_factory.h - Operation Registry
main/ud_functions/udf_manager.h - Distributed Operations
main/operations/reduce_op.h - Storage Interface
main/storage/storage.h - Memory Storage
main/storage/memory_storage.h - NVMe Storage
main/storage/nvme_storage.h - GPU Scheduler
main/gpu/scheduler.h - XLA Kernel Base
main/ud_functions/xla_kernel_base.h
Other components
- Client
Connections 28
- Client → Node CLI: submits commands
- Node CLI → Compute Node: uses node
applications/node/node_cli.cc - Node CLI → Tensor Factory: uses factory
applications/node/node_cli.cc - Matrix Workloads → Parsed Commands: builds commands
applications/bmm/generate_bmm.cc - Feed-Forward Workloads → Two-Layer Compiler: uses compiler
applications/ffnn/generate_ffnn.cc - Feed-Forward Workloads → Parsed Commands: uses command model
applications/ffnn/generate_ffnn.cc - Feed-Forward Workloads → Tensor Model: uses tensors
applications/ffnn/generate_ffnn.cc - XLA Kernel App → Operation Registry: registers kernels
applications/xla_kernel/main.cc - XLA Kernel App → XLA Kernel Base: uses base
applications/xla_kernel/main.cc - Parsed Commands → Two-Layer Compiler: compiled by
- Two-Layer Compiler → Execution Commands: generates commands
main/commands/two_layer_compiler.h - Two-Layer Compiler → Tensor Model: plans tensor placement
- Command Runner → Reservation Station: dispatches work
main/commands/command_runner.h - Command Runner → Operation Registry: uses registry
main/commands/command_runner.h - Command Runner → MPI Communication: uses communication
main/commands/command_runner.h - Reservation Station → Command Handlers: dispatches commands
main/commands/reservation_station.h - Command Handlers → Execution Commands: handles commands
main/commands/command_handlers/command_handler_apply.cc - Coordinator → MPI Communication: coordinates nodes
- Compute Node → Command Runner: runs commands
- Distributed Operations → MPI Communication: communicates tensors
main/operations/reduce_op.h - Distributed Operations → Storage Interface: accesses tensors
main/operations/reduce_op.h - Distributed Operations → Operation Registry: selects operation
main/operations/reduce_op.h - Storage Interface → Memory Storage: selects backend
main/storage/storage.h - Storage Interface → NVMe Storage: selects backend
main/storage/storage.h - Tensor Factory → Tensor Model: creates tensors
main/tensor/tensor_factory.h - Operation Registry → Tensor Factory: resolves tensor types
main/ud_functions/udf_manager.h - Operation Registry → GPU Scheduler: uses scheduler
main/ud_functions/udf_manager.h - XLA Kernel Base → Operation Registry: registered through