Programming Languages & Software Engineering

Programming Languages & Software Engineering

CSE373: Data Structures & Algorithms Lecture 12: Amortized Analysis and Memory Locality Lauren Milne Spring 2015 Announcements Homework 3 due on Wednesday at 11pm Catie back Monday Spring 2015 CSE 373 Data structures and Algorithms 2 Amortized Analysis In amortized analysis, the time required to perform a sequence of data structure operations is averaged over all the operations performed.

Typically used to show that the average cost of an operation is small for a sequence of operations, even though a single operation can cost a lot Spring 2015 CSE 373 Data structures and Algorithms 3 Amortized Analysis Recall our plain-old stack implemented as an array that doubles its size if it runs out of room Can we claim push is O(1) time if resizing is O(n) time? We cant, but we can claim its an O(1) amortized operation Spring 2015 CSE 373 Data structures and Algorithms

4 Amortized Complexity We get an upperbound T(n) on the total time of a sequence of n operations. The average time per operation is then T(n)/n, which is also the amortized time per operation. If a sequence of n operations takes O(n f(n)) time, we say the amortized runtime is O(f(n)) If n operations take O(n), what is amortized time per operation? O(1) per operation If n operations take O(n3), what is amortized time per operation? O(n2) per operation The worst case time for an operation can be larger than f(n), but amortized guarantee ensures the average time per operation for any sequence is O(f(n)) Spring 2015 CSE 373 Data structures and Algorithms

5 Building Up Credit Can think of preceding cheap operations as building up credit that can be used to pay for later expensive operations Because any sequence of operations must be under the bound, enough cheap operations must come first Else a prefix of the sequence would violate the bound Spring 2015 CSE 373 Data structures and Algorithms 6 Example #1: Resizing stack A stack implemented with an array where we double the size of the array if it becomes full Claim: Any sequence of push/pop/isEmpty is amortized O(1) per

operation Need to show any sequence of M operations takes time O(M) Recall the non-resizing work is O(M) (i.e., M*O(1)) The resizing work is proportional to the total number of element copies we do for the resizing So it suffices to show that: After M operations, we have done < 2M total element copies (So average number of copies per operation is bounded by a constant) Spring 2015 CSE 373 Data structures and Algorithms 7 Amount of copying Claim: after M operations, we have done < 2M total element copies Let n be the size of the array after M operations Then we have done a total of:

n/2 + n/4 + n/8 + INITIAL_SIZE < n element copies Because we must have done at least enough push operations to cause resizing up to size n: M n/2 So 2M n > number of element copies Spring 2015 CSE 373 Data structures and Algorithms 8 Other approaches If array grows by a constant amount (say 1000), operations are not amortized O(1) After O(M) operations, you may have done (M2) copies If array shrinks when 1/2 empty, operations are not amortized O(1) Terrible case: pop once and shrink, push once and grow, pop once and shrink,

If array shrinks when 3/4 empty, it is amortized O(1) Proof is more complicated, but basic idea remains: by the time an expensive operation occurs, many cheap ones occurred Spring 2015 CSE 373 Data structures and Algorithms 9 Example #2: Queue with two stacks A clever and simple queue implementation using only stacks class Queue { Stack in = new Stack(); Stack out = new Stack(); void enqueue(E x){ in.push(x); } E dequeue(){ if(out.isEmpty()) { while(!in.isEmpty()) { out.push(in.pop());

} } return out.pop(); } } Spring 2015 CSE 373 Data structures and Algorithms enqueue: A, B, C C B A in out 10

Example #2: Queue with two stacks A clever and simple queue implementation using only stacks class Queue { Stack in = new Stack(); Stack out = new Stack(); void enqueue(E x){ in.push(x); } E dequeue(){ if(out.isEmpty()) { while(!in.isEmpty()) { out.push(in.pop()); } } return out.pop(); } } Spring 2015 CSE 373 Data structures and Algorithms dequeue

A B C in out 11 Example #2: Queue with two stacks A clever and simple queue implementation using only stacks class Queue { Stack in = new Stack(); Stack out = new Stack(); void enqueue(E x){ in.push(x); } E dequeue(){ if(out.isEmpty()) { while(!in.isEmpty()) { out.push(in.pop());

} } return out.pop(); } } Spring 2015 CSE 373 Data structures and Algorithms enqueue D, E A E D B C in

out 12 Example #2: Queue with two stacks A clever and simple queue implementation using only stacks class Queue { Stack in = new Stack(); Stack out = new Stack(); void enqueue(E x){ in.push(x); } E dequeue(){ if(out.isEmpty()) { while(!in.isEmpty()) { out.push(in.pop()); } } return out.pop(); } } Spring 2015

CSE 373 Data structures and Algorithms dequeue twice CBA E D in out 13 Example #2: Queue with two stacks A clever and simple queue implementation using only stacks class Queue { Stack in = new Stack(); Stack out = new Stack(); void enqueue(E x){ in.push(x); }

E dequeue(){ if(out.isEmpty()) { while(!in.isEmpty()) { out.push(in.pop()); } } return out.pop(); } } Spring 2015 CSE 373 Data structures and Algorithms dequeue again DCBA E in out

14 Analysis dequeue is not O(1) worst-case because out might be empty and in may have lots of items But if the stack operations are (amortized) O(1), then any sequence of queue operations is amortized O(1) The total amount of work done per element is 1 push onto in, 1 pop off of in, 1 push onto out, 1 pop off of out When you reverse n elements, there were n earlier O(1) enqueue operations to average with Spring 2015 CSE 373 Data structures and Algorithms 15 When is Amortized Analysis Useful?

When the average per operation is all we care about (i.e., sum over all operations), amortized is perfectly fine If we need every operation to finish quickly (e.g., in a web server), amortized bounds may be too weak Spring 2015 CSE 373 Data structures and Algorithms 16 Not always so simple Proofs for amortized bounds can be much more complicated Example: Splay trees are dictionaries with amortized O(log n) operations See Chapter 4.5 if curious For more complicated examples, the proofs need much more sophisticated invariants and potential functions to describe how earlier cheap operations build up energy or money to pay for later expensive operations

See Chapter 11 if curious But complicated proofs have nothing to do with the code (which may be easy!) Spring 2015 CSE 373 Data structures and Algorithms 17 Switching gears Memory hierarchy/locality Spring 2015 CSE 373 Data structures and Algorithms 18 Why do we need to know about the memory hierarchy/locality?

One of the assumptions that Big-O makes is that all operations take the same amount of time Is this really true? Spring 2015 CSE 373 Data structures and Algorithms 19 Definitions A cycle (for our purposes) is the time it takes to execute a single simple instruction (e.g. adding two registers together) Memory latency is the time it takes to access memory Spring 2015 CSE 373 Data structures and Algorithms

20 Time to access: ~16-64+ registers SRAM CPU Cache 1 ns per instruction 2-10 ns 8 KB - 4 MB Main Memory DRAM

40-100 ns 2-10 GB Disk many GB Spring 2015 a few milliseconds (5-10 million ns) CSE 373 Data structures and Algorithms 21 What does this mean? It is much faster to do: 5 million arithmetic ops 2500 L2 cache accesses

Than: 1 disk access 1 disk access 400 main memory accesses 1 disk access Why are computers build this way? Physical realities (speed of light, closeness to CPU) Cost (price per byte of different storage technologies) Under the right circumstances, this kind of hierarchy can simulate storage with access time of highest (fastest) level and size of lowest (largest) level Spring 2015 CSE 373 Data structures and Algorithms 22

Spring 2015 CSE 373 Data structures and Algorithms 23 Processor-Memory Performance Gap Spring 2015 CSE 373 Data structures and Algorithms 24 What can be done? Goal: attempt to reduce the accesses to slower levels Spring 2015

CSE 373 Data structures and Algorithms 25 So, what can we do? The hardware automatically moves data from main memory into the caches for you Replacing items already there Algorithms are much faster if data fits in cache (often does) Disk accesses are done by software (e.g. ask operating system to open a file or database to access some records) So most code just runs, but sometimes its worth designing algorithms / data structures with knowledge of memory hierarchy To do this, we need to understand locality Spring 2015 CSE 373 Data structures and Algorithms 26

Locality Temporal Locality (locality in time) If an item (a location in memory) is referenced, that same location will tend to be referenced again soon. Spatial Locality (locality in space) If an item is referenced, items whose addresses are close by tend to be referenced soon. Spring 2015 CSE 373 Data structures and Algorithms 27 How does data move up the hierarchy? Moving data up the hierarchy is slow because of latency (think distance to travel) Since were making the trip anyway, might as well carpool Get a block of data in the same time we could get a byte

Sends nearby memory because Spatial Locality Its easy Likely to be asked for soon (think fields/arrays) Once a value is in cache, may as well keep it around for a while; accessed once, a value is more likely to be accessed again in the near future (as opposed to some random other value) Temporal Locality Spring 2015 CSE 373 Data structures and Algorithms 28 Cache Facts Definitions: Cache hit address requested is in the cache Cache miss address requested is NOT in the cache Block or page size the number of contiguous

bytes moved from disk to memory Cache line size the number of contiguous bytes moved from memory to cache Spring 2015 CSE 373 Data structures and Algorithms 29 Examples x = a + 6 miss y = a + 5 hit z = 8 * a hit Spring 2015 CSE 373 Data structures and Algorithms 30

Examples x = a + 6 miss x = a[0] + 6 miss y = a + 5 hit y = a[1] + 5 hit z = 8 * a hit z = 8 * a[2] hit Spring 2015 CSE 373 Data structures and Algorithms 31

Examples x = a + 6 miss x = a[0] + 6 miss y = a + 5 hit y = a[1] + 5 hit z = 8 * a hit z = 8 * a[2] hit temporal locality Spring 2015 spatial locality CSE 373 Data structures and Algorithms

32 Locality and Data Structures Which has (at least the potential) for better spatial locality, arrays or linked lists? 100 101 102 103 104 105 106 1 2 3 4 5 6

7 a[0] a[1] a[2] a[3] a[4] a[5] a[6] Spring 2015 CSE 373 Data structures and Algorithms 33 Locality and Data Structures Which has (at least the potential) for better spatial locality, arrays or linked lists? e.g. traversing elements 100 101 102 103 104 105 106 1

2 3 4 5 6 7 a[0] a[1] a[2] a[3] a[4] a[5] a[6] miss hit hit cache line size

hit miss hit hit cache line size Only miss on first item in a cache line Spring 2015 CSE 373 Data structures and Algorithms 34 Locality and Data Structures Which has (at least the potential) for better spatial locality, arrays or linked lists?

e.g. traversing elements 100 101 1 Spring 2015 300 301 2 50 51 3 CSE 373 Data structures and Algorithms 62

63 4 35 Locality and Data Structures Which has (at least the potential) for better spatial locality, arrays or linked lists? e.g. traversing elements 100 101 1 miss hit 300 301 2 miss hit

50 51 3 miss hit 62 63 4 miss hit Miss on every item (unless more than one randomly happen to be in the same cache line) Spring 2015

CSE 373 Data structures and Algorithms 36

Recently Viewed Presentations

  • Market Positioning

    Market Positioning

    Step 4: Add Digital Capability to the Magnum. There are numerous installations of hybrid production lines . currently in operation. There a 9 Goss commercial presses using hybrid printing to produce: Catalogues. Direct Mail. Inserts. Commercial
  • tiate and Develop Industry Relationship Andrew Marsh Hothouse

    tiate and Develop Industry Relationship Andrew Marsh Hothouse

    Andrew Marsh. Collaboration - Why? Your Motivations? ... Focus on building deeper relationships with mutual advantage. Where to focus. Conferences. Online forum. LinkedIn groups. Industry showcase. Student project showcase. IRDG events. How will I present myself.
  • Governance Ecosystems CSR in the Latin American Mining

    Governance Ecosystems CSR in the Latin American Mining

    Governance Ecosystems CSR in the Latin American Mining Sector. Premise 1:Poverty, social exclusion, and environmental degradation, while often clearly exacerbated by irresponsible business, exist within broader local and global political economic context in which irresponsibility and lack of accountability are...
  • Self Organizing Maps

    Self Organizing Maps

    A histogram is formed based on the hits on WCM. Use this histogram as fingerprint for DCM. References [1] T.Honkela, S.Kaski, K.Lagus,T.Kohonen. WEBSOM- Self Organizing Maps of Document Collection. (1997) [2] T.Honkela, S.Kaski, K.Lagus,T.Kohonen. Exploration of full text Databases with...
  • Montague Lord M., Lady M., Romeo, Benvolio (Romeo&#x27;s Cousin)

    Montague Lord M., Lady M., Romeo, Benvolio (Romeo's Cousin)

    Capulet asks him to wait 2 years to marry Juliet at first, but later tries to force Juliet to marry him immediately. Rosaline. She is a beautiful girl who Romeo loves at the beginning of the play. She has made...
  • AUTO PSRU&#x27;s LLC FORUM

    AUTO PSRU's LLC FORUM

    Need a gear ratio to get from Propeller to Engine RPM's. For GM LS V8 engines the BW350 gear ratio is 1.667:1. RPM limit for stock GM crate engines is 6000 RPM. Safe continuous RPM at 75% of limit is...
  • Middle Fossa Approach to the IAM

    Middle Fossa Approach to the IAM

    objectives. give a differential diagnosis for the symptoms of. dysphonia. dysphagia . odynophagia . define the NICE guidelines for hoarseness. list the differential diagnosis for a nasal condition resulting in nasal obstruction with other symptoms such as bleeding, epiphora and...
  • Urban game Pre - game set up

    Urban game Pre - game set up

    Urban game Pre - game set up ... a number of new mechanical inventions for farms are developed. One is called the seed drill and another is the horse-drawn cultivator. Also, farmers begin to experiment with new, more productive farming...