BORPH

1. BORPH Presented by Brandon Hamilton

2. Reconfigurable Computing BEE2 Compute Module ROACH

3. Introduction • Operating system support for reconfigurable computers • Isolation of low level details of FPGA design • Remove need for device driver concerns during application development • Balance between flexibility and genericity, and high performance.

4. BORPH • Berkeley Operating system for ReProgrammable Hardware (BORPH) • PhD Thesis Fall 2007 University of California, Berkeley • Hayden Kwok-Hay So and Robert W. Brodersen • Running on BEE2 at Berkeley Wireless Research Center

5. BORPH overview • Extension of Linux kernel • Implements UNIX kernel semantics • Implementation of “Hardware processes” • Runtime general filesystem support • Interprocess communication (IPC) between software and hardware processes • Manages Reconfigurable resources

6. Hardware Processes • Logical separation of reconfigurable fabric and underlying hardware support platform • Reconfigurable Hardware region (HWR) is smallest unit of reconfigurable fabric. (e.g. an FPGA) • A hardware process is an executing instance of a gateware program • Independent execution (not under control of software) • Eliminates the need to write software to control gateware programs • Views reconfigurable fabrics as computing resources

7. Extension of Linux kernel • BOF (BORPH Object file) - executable binary file format contains information such as configuration of reconfigurable fabrics. • Able to be executed by any normal UNIX process creation techniques (fork, exec). • Hardware process has own logical virtual memory space.

8. Implements UNIX kernel semantics • Preserves UNIX process creation semantics (parent-child relationships, process group info) as well as signals (SIGTERM, SIGKILL etc.) • Process is run in 'Interruptable sleep' state so that it does not get scheduled to run on the processor.

9. Hardware Processes

10. BORPH Kernel

11. Interprocess Communication • Currently no shared memory between hardware process and other processes. • Message passing OR data streaming modes. • Kernel creates connection to gateware via 2 unidirectional IO connection points (analogous to file descriptors)

12. IO Connection Points

13. System calls • In standard software kernels, system calls (API into the kernel) are synchronous (blocking) • BORPH hardware system call interface -asynchronous and parallel • User programs use request-acknowledge protocol for hardware system calls.

14. Access to reconfigurable resources • IOREG interface used to access hardware registers, FIFO's, on/off-chip memories from software • Creates virtual files corresponding to hardware constructs in gateware design - /proc/<pid>/hw

15. Tested Performance • Bottleneck is communication performance between control and user FPGAs • Use of DMA greatly improved performance. Process Creation File Streaming

16. Conclusions • Allows gateware designs to play active role – p2p • Abstraction model / interface • Future considerations • Semantics (blocking, swapping, parallel I/O) • Gateware debugging capabilities • Scalability • Optimization

17. Reference • Hayden Kwok-Hay So and Robert W. Brodersen, “BORPH: An Operating System for FPGA-Based Reconfigurable Computers”, EECS Department, University of California, Berkeley, July 2007