Parallel computing is a style of programming where a computation is broken into parts that run simultaneously across multiple processors, cores, or machines. The motivation is straightforward: a single core has a clock speed ceiling, and modern CPUs gain performance by adding more cores rather than running each core faster. To take advantage of that, programs have to be written with parallelism in mind.
Not every program benefits equally. Amdahl's law shows that the sequential fraction of a program — the part that cannot be parallelized — sets a hard ceiling on speedup regardless of how many cores you add. Gustafson's law is the more optimistic counterpart: if you scale the problem size alongside the hardware, speedup grows linearly. In practice, HPC workloads follow Gustafson's regime — you buy more nodes to solve a bigger problem, not just to solve the same one faster.
Parallel computing splits into three broad paradigms based on where the parallelism lives.
Shared-memory parallelism runs multiple threads on a single machine with a common address space. OpenMP is the standard approach in C, C++, and Fortran: a few #pragma omp directives turn a serial loop into a parallel one. Threads communicate by reading and writing shared variables, which makes synchronization — mutexes, barriers, atomics — the main source of bugs and overhead.
Distributed-memory parallelism runs processes across separate machines (or separate address spaces on one machine), each with its own private memory. MPI is the dominant standard. Processes communicate explicitly by sending and receiving messages. There is no shared state to race on, but the programmer is responsible for every byte that crosses a process boundary. MPI is the backbone of large cluster workloads.
GPU parallelism offloads computation to a GPU, which can run thousands of lightweight threads simultaneously. CUDA is NVIDIA's programming model for this. GPU parallelism is best suited for problems where the same operation is applied to a large array of data — matrix multiplication, FFTs, stencil operations. The bottleneck is usually memory bandwidth and the cost of transferring data between host (CPU) memory and device (GPU) memory.
Parallel programs introduce failure modes that serial programs don't have: race conditions, deadlocks, false sharing, memory ordering issues. A race condition occurs when two threads read and write shared data without synchronization and the outcome depends on the order of execution. A deadlock occurs when two threads are each waiting for a lock the other holds. False sharing is a subtler hardware-level issue: two threads write to different variables that happen to sit in the same cache line, causing the cache coherence protocol to thrash.
On the performance side, the roofline model is a useful frame for understanding whether a kernel is compute-bound or memory-bandwidth-bound, which determines where to focus optimization effort. For quick empirical benchmarks, SAXPY — a simple vector operation — is a standard starting point for measuring memory bandwidth.