19-Task-Parallelism

Question 1

What is task parallelism?

Answer 1

Executing different tasks on different processors, not just data parallel operations

Question 2

What is the difference from data parallelism?

Answer 2

Data parallelism: same operation on different data; Task parallelism: different operations

Question 3

What is latency hiding?

Answer 3

Overlapping communication with computation to hide network delays

Question 4

What is blocking vs non-blocking?

Answer 4

Blocking waits for completion, non-blocking returns immediately

Question 5

What is MPI_Wait()?

Answer 5

Blocks until non-blocking operation completes

Question 6

What is MPI_Test()?

Answer 6

Non-blocking check if operation completed

Question 7

What are command queues?

Answer 7

GPU queues for submitting work asynchronously

Question 8

What are events?

Answer 8

Track dependencies between GPU operations

Question 9

What is event-based synchronization?

Answer 9

Use events to chain operations: kernel depends on data transfer

Question 10

What is the work-span model?

Answer 10

Estimates maximum parallel speedup based on work (total operations) and span (longest dependency chain)

Question 11

What is work in work-span model?

Answer 11

Total operations needed (serial time)

Question 12

What is span in work-span model?

Answer 12

Time on infinite processors (longest dependency chain)

Question 13

What is the work-span speedup limit?

Answer 13

S ≤ work/span

Question 14

What are superscalar sequences?

Answer 14

Runtime schedules tasks based on dependencies, not explicit synchronization

Question 15

What is the advantage of task graphs?

Answer 15

Automatic optimization, cleaner code, runtime handles synchronization

Question 16

What is the disadvantage?

Answer 16

Less control over execution order

Question 17

What is OpenCL profiling?

Answer 17

Measuring kernel execution time using events

Question 18

What is CL_QUEUE_PROFILING_ENABLE?

Answer 18

Enables timing measurements on command queue

Question 19

What is clGetEventProfilingInfo()?

Answer 19

Retrieves start/end times for profiling

Question 20

What is the timing calculation?

Answer 20

(end - start) * 1e-9 seconds

Question 21

What is multiple command queue benefit?

Answer 21

Overlap different operations across queues

Question 22

What is the task graph representation?

Answer 22

Directed acyclic graph: nodes=tasks, edges=dependencies

Question 23

What is critical path?

Answer 23

Longest path in task graph, determines minimum execution time