Kang-Pin Chan

Carnegie Mellon University

University of Southern California

National Cheng Kung University

Overall GPA: 3.92/4.3 (Major GPA: 4.22/4.3)

Relevant Course: DATA STRUCTURE (97, A+), ALGORITHMS (93, A+), OPERATING SYSTEMS (89, A)

Intel

• Managed Windows Graphic Device Driver flow with C/C++ to acquire BSOD or TDR data from memory and process data to make WinDbg Tool’s features available in both live and dump file environments

• Constructed Python script to auto generate 300+ lines of code from driver functions during compile state to reduce effort of maintenance between tool and driver

• Solved Unit Testing Framework issues with C++ enabling users to reproduce driver scenario captured in ETL files

• Proposed and fix problems about DiAna tool that uses C# code to parse ETL files into .xml processed by UTF tool

• Tools were used by 50+ debugging engineers in four global teams for log processing and core data retrieval

Robotics Lab - National Cheng Kung University

• Built an automated guided vehicle (AGV) capable of automatically avoiding people and following paths based on Lidar detection in order to accomplish commands called by the User.

• Collected map data by test driving the AGV through school buildings and saved data for the Robot Operating System module to assist with navigation to the destination.

• Employed the FER-2013 dataset to build an emotion detection model that can detect 7 kinds of emotion, and integrated the model with a RealSense camera to collect real-time feedback of the user’s reactions so the AGV could modify its service in dynamic response to the user’s feelings.

• Applied Mediapipe to detect the position of the palm and recognize critical points on the hand to identify and match gestures for 1-5, enabling the user to provide a final rating after each service.

/dev/mem research

• Utilized Linux tool Crash to validate the memory mapping mechanisms regarding the /DEV/MEM image file.

• Reserved physical memory to avoid mapping by the memory management(MMU) unit and used mmap to modify the specified reserved memory.

• Validated the structure of Linux Five-level page tables by calculating the indices of PGD, PMD, PUD, and PTE and used Crash to confirm the output.

• Circumvented the kernel-protection mechanism — KPTI, ASLR, Spectre, and KASLR — to acquire a fixed virtual base address and achieve an ideal outcome.

• Computed the offset of attributes in the task struct to modify the information in the process and traversed all processes in the system.

Fibdrv

• ExploitedLinux virtual file system (VFS) interface to operate device driver as a file descriptor.

• Implemented a character device driver to calculate i‑th fibonacci number to response a read from user space.

• Supported in‑kernel big number arithmetic allows calculation with arbitrary integer number.

• Exploited sysfs to get the calculation time to avoid the overhead of kprintf.

Linked-list sort

• Replaced stdlib rand() with Pseudo Random Number Generator (PRNG) - MT19937, CRYPTMT to establish test data for sorting.

• Implemented non-recursive quicksort for linked list — 10% improvement in speed against the recursive method.

• Built an efficient mergesort for linked list — saved 15% of heap usage against naive implementation.

• Extracted code from lib/list sort.c to build a standalone program and applied several linux tools to compare the performance and dynamic stack usage between the quicksort and lib/list sort.c.

Line-bot-elder

• Used the concept of finite state machine to implement the project.

• Used web crawler to catch image and data from wikipedia and google, also process JSON file to get train schedule.

Key-Value-Storage

• Implemented by C++ and able to handle over 16GB size of data.

• Use LRU as data organization strategy and apply Copy-on Write strategy to reduce the output I/O effort.

ETL-Processing

• Implemented by C++ and able to handle at least 1GB size of data.

• Multi-threaded programming - up to 40\% speedup against single-threaded implementation.

• Compared and analysed the different of thread's number can influence the hardware activity.

External-Sort Benchmark & Improvement

• Constructed the External-Sort code in C++ and was able to handle over 16GB size of data, which outstrips the limit of main memory.

• Used K-way-external-sort to improve the sort’s cache locality up to 52% and obtained a 23% increase in speed against the 2-way-external-sort.

• Changed the output strategy to write-back by creating an output buffer and export the sorting outcome in scale which obtained a 20% increase in speed against naive implementation.

• Used the linux tool (perf) to measure the hardware activity and observe the performance bottleneck.