Final Lecture Review

Question 1

How to calculate total time and avg time to complete order for boss-worker:
ex)
6 workers 11 toy orders
120ms per toy order

Answer 1

total time for 5 workers
5 order completed -> 120ms
next 5 completed -> 120ms
last (11th) order completed -> 120ms 
= 360ms 

avg time = 5(120) +5(240) + 360 / 11 = 2160/ 11 = 196ms

Question 2

How to calculate total time and avg time to complete order for pipeline:
ex)
6 workers 11 toy orders
20ms per pipeline stage

Answer 2

total time for 1 order -> 120ms
offset is 20ms

10 remaining orders after first
total time = 120 + 20(10) = 320ms

avg time = 120 + 140 + 160 … + 320 / 11 = 2420 / 11 = 220ms

Question 3

Why are threads useful?

Answer 3

Parallelization speed up
Specialization -> leads to hot cache
Lower memory requirement than processes and cheaper synchronization

Usefulness depends on:

• Depends on the situation and the metrics we use
• Depends on the workload

Question 4

Define once again the tradeoffs between MP and MT

Answer 4

MP benefits: 
- easier to develop 
- relies on the OS 
MP costs: 
- more costly memory-wise due to expensive context switching and IPC 
- tricky port setup 
- costly to maintain shared state 
MT benefits: 
- shared address space 
- little cost of context switching 
MT costs: 
- very hard to implement 
- synchronization is difficult to handle

Question 5

How does the event-driven model work?

Answer 5

Event driven model:
- Process request until wait is necessary
- Then switch to another request when waiting on a request
- This continues until all requests are served
- Single address space and single flow of control
- Single process and single thread (helper processes though as needed)
- Smaller memory requirement and no context switching
No synchronization

Compared to MT - context switching to other threads wastes time that could have been dedicated to processing other requests

Question 6

Of the three models, which of the following requires the least amount of memory?

Boss-worker model
Pipeline model
Event-driven model

Answer 6

Event-driven model

Extra memory is only required for the helper thread in event-driven because of the concurrent blocking I/O, not for all concurrent requests

Question 7

Experiments and the winners from the FLASH paper

Answer 7

Single file test (synthetic) -> SPED (greater reliance on cache and didn’t need to check cache for memory presence)

Owlnet trace -> AMPED barely beat out sped due to blocking I/O advantage (helpers); SPED occasionally blocked

CS trace (Larger trace which requires I/O and not fed into cache ) -> AMPED won due to smaller memory footprint and more memory for caching; SPED performed worst because of lacking async I/O

Optimizations connection rate -> all optimizations performed best and no optimizations performed the worst (optimizations are very important)

Performance under WAN -> performance improvements when more clients were added