HPC Warmup Quiz

The industry standard for programming for NVIDIA GPUs as a accelerator is

• OpenMP

• MPI

• GCC

• CUDA

The industry standard for threaded multi-core programming in HPC is

• OpenMP

• CUDA

• MPI

• GCC

The industry standard protocal for distributed computing in HPC is

• CUDA

• GCC

• OpenMP

• MPI

In threaded parallelism, when multiple threads can potentially write to the same memory location at the same time, this is called ...

• False Sharing

• Page Fault

• Race Condition

• Traffic Jam

When two or more threads modify different variables that happen to share a cache line, causing unnecessary invalidation and slowdown, this is known as…

• Memory thrashing

• False sharing

• A lock convoy

• Load imbalance

Amdahl’s Law tells us that speedup is limited by…

• The number of MPI ranks

• The power consumption of the node

• How fast the interconnect is

• The fraction of the code that is sequential

A job with --nodes=1 --ntasks=16 will:

• Use 16 nodes

• Launch 16 tasks on the same node

• Spread 16 tasks across 16 nodes

• Use 16 CPU threads but 1 MPI rank

For a program that is 95% parallelizable, a reasonable choice for the number of nodes to use is:

• 4 nodes

• 8 nodes

• 20 nodes

• 1000 nodes

The biggest difference between distributed and threaded parallelism is:

• Distributed parallelism always runs faster

• Threaded parallelism requires an interconnect network

• Distributed programs communicate by messages; threaded programs share memory

• Threaded programs use MPI and distributed programs use OpenMP

In OpenMP, which variable is always private by default inside parallel for loops?

• The loop index

• Any array

• All global variables

• Anything declared before the loop