L20 & L21

Question 1

How has performance optimisations, such as OOE, affected the correctness of a program?

Answer 1

Loads and stores that are issued out of order reach memory out of order. Therefore, the memory becomes inconsitent across cores.

Question 2

What causes memory inconsistency in optimised code?

Answer 2

out-of-order stores

Question 3

What are the 3 components of a memory consistency model?

Answer 3

 Clear definition of how stores become visible
 Clear definition of what is allowed and not
 Clear context for HW and SW optimizations

Question 4

Explain what the release consistency model is.

Answer 4

Release consistency: contract between programmer and system
* Shared access is protected by synchronisation
* Critical section is assumed atomic
* Operation order does not have external effects

Question 5

What are the 2 types of memory fence/barrier?

Answer 5

Full memory fence: all memory operations are complete when the fence completes

Store memory fence does the same but only for stores

Question 6

What are the 2 different ways to code a memory fence/barrier

Answer 6

DMB = data memory barrier

LDAR = load-acquire / STLR = store-release

Question 7

What are the differences between coherence and consistency?

Answer 7

Coherence:
 Single memory location
 All cores see the same updates
 Not interested in the relative ordering of different updates
 Done automatically by hardware

Consistency:
 About the overall system state
 Not interested in keeping data coherent
 A contract between programmer, compiler, and hardware

Question 8

What is a way to enforce partial ordering?

Answer 8

Memory fence/barrier

Question 9

What is functional programming? What is an important property of it?

Answer 9

Functional programming: transforms inputs to outputs by applying a set of functions

Referential transparency is an important property of functional programming
 Can replace a name with its value in the program
 No updateable state

Question 10

What is a higher level language that can be used to translate program to data flow representation?

Answer 10

o SISAL: stream and iteration in a single assignment language

Question 11

What is the data flow representation components?

Answer 11

 Token queue: buffering + inserting new data
 Matching store: tokens matched with instructions
 Instruction store: coding of the dataflow graph
 Processor bank: a number of processors

Question 12

How can the dataflow representation be represented as a graph?

Answer 12

 Instructions = nodes
 Data flow = edges

Question 13

What are examples of pure (“coarse-grained”) data flow?

Answer 13

Task-based computing
OpenMP tasking
OpenStream