Parallel Programming & Threading

OpenMP Medicaid Data Summarizer (A0.2)

Parallelized a Medicaid CSV summarizer that aggregates total payments by HCPCS code, comparing a sequential baseline against an OpenMP implementation with per-thread local hash maps.

• • Reads the entire CSV into memory, then splits work across OpenMP threads by file chunk
• • Aligns chunk boundaries to full line breaks so threads never parse partial rows
• • Each thread accumulates totals in a local unordered_map, merged into a global result at the end
• • Custom Stopwatch instrumentation measures read, summarize, copy, and sort phases separately

MPI Distributed Web Crawl Analysis (A2)

Distributed analysis of a large web crawl dataset across 16 MPI ranks on an HPC cluster, computing global statistics and top/bottom URL rankings.

• • Round-robin file assignment distributes 25,440 crawl files across MPI ranks
• • MPI_Reduce and MPI_Allreduce for global sums, counts, min/max, and distributed median
• • Distributed median via binary search with MPI_Allreduce count aggregation per iteration
• • MPI_Gather collects per-rank top/bottom candidates for global ranking merge on rank 0

pthread Parallel Image File Writes (a3.5_threading)

Parallelized image file writes by splitting data into blocks and dispatching each block to a POSIX thread, with mutex-protected shared file I/O.

• • Splits image data into fixed-size blocks and spawns a pthread worker per block
• • pthread_mutex_t guards fseek/fwrite calls in the shared write delegate
• • Bounded thread pool pattern — joins threads when MAX_THREADS is reached before spawning more
• • Each thread receives its own ThreadArgs struct with block offset, size, and block number

CPU Process Scheduling Simulators (a3_proc_scheduling)

Implemented four CPU scheduling algorithms that simulate how an OS manages concurrent processes on a single core, computing waiting time and turnaround time metrics.

• • First-Come First-Served (FCFS) with non-preemptive sequential execution
• • Shortest Job First (SJF) selecting the shortest available burst at each decision point
• • Round Robin with configurable time quantum and queue rotation on completion
• • Shortest Remaining Time First (SRTF) with implicit preemption each clock tick

Performance at Scale

The MPI web crawl analysis processed over 658 million rows from 25,440 files in approximately 323 seconds across 16 MPI ranks on an HPC cluster. The sequential Medicaid CSV summarizer took roughly 438 seconds for the read-and-summarize phase alone, motivating the OpenMP parallelization with thread-local hash maps and line-boundary-safe chunking.

Concurrency Lessons

These assignments reinforced the tradeoffs between parallel paradigms: thread-local accumulation avoids lock contention but requires a merge step, mutex-protected I/O serializes writes but keeps file state consistent, and scheduling algorithm choice (SRTF vs Round Robin) directly impacts waiting time and responsiveness under different workload patterns.