Research On Hardware Support Profiling Of Dynamic Binary Translation

Profiling involves collecting run-time information of the system. You can use the profile information to optimize the system. A good profiling system can obtain very large improvement on performance for program or overall system. Moreover, profiling system using Hardware/Software co-design technique incurs lower overhead, and obtains high performance. The HW/SW co-design technique changes the architecture of Dynamic Binary Translation (DBT) based on software-only methods, and can obtain the merits of both hardware and software to achieve better overall system performances compared the software-only virtual machines. Because the hardware can accelerate virtual machine systems to overcome the performance bottleneck, the overall performance of profiling system based on DBT can increase. Also, it is transparent to users.The traditional profiling strategy in dynamic binary translator is based on the basic block, the edge jump and the path tracing. In general, the method using software-only profiling typically incurs about 30% overhead of overall system. By hardware-support, dynamic optimization system can improve overall performance prominently. Moreover, the difficulty in HW/SW co-design is the communication overheads and partition between hardware and software. To the optimization process of DBT, This thesis replaces the traditional software-only profiling by hardware-support profiling technique to reduce system overhead and to improve the overall performance of DBT system. We can obtain the better performance by reducing communication overhead to the minimum. We design and implement our idea on Dynamic Binary Translator, CrossBit. The main contents of this thesis include: the hardware-software partition of the whole CrossBit, how to communicate efficiently between hardware and software, and how to collect efficiently profile information. The thesis proposes a novel method to reduce overall overhead and to improve performance of DBT systems by hardware-support.In this thesis, we implement hardware-support profiling system on CrossBit platform, and the main design and implementation concentrate on the overhead and accuracy of collect profile. This design is verified by the SPEC CPU 2000 program, Experimental data show that the overhead of CrossBit increases about 2.7%, which means that we can collect correct profile information on binary translator by low overhead.