1. Compact Data Structures and Applications Gil Einziger and Roy Friedman Technion, Haifa

2. Approximate Set Membership The problem: The problem: – Maintain enough state to approximately answer set membership queries. (add/query). Things to consider: Things to consider: – False positive probability. – No false negatives! – A tradeoff between space and the false positive probability. Bloom filters are the classical example.

3. Application Example: Black List Your Computer Client: Client: Web browser Server: Server:Let’s say Google’s data center. Mission: Mission:Check if each accessed URL is part of the black list. Challenge: Challenge: The black list is too big to fit in memory. Data Center Trivial Solution: access data center for each URL Approximate Set: Access data center only in case of a positive answer. (must access it as the answer may be wrong) But… there are no false negatives! Therefore every site that tests negative is not on the list and we do not have to contact the datacenter.

4. Bloom Filters An array BF of m bits and k hash functions {h 1,…,h k } over the domain [0,…,m-1] Adding an object obj to the Bloom filter is done by computing h 1 (obj),…, h k (obj) and setting the corresponding bits in the Bloom filter. Checking for set membership for an object cand is done by computing h 1 (cand),…, h k (cand) and verifying that all corresponding bits are set m=11, k=3, 111 h 1 (o1)=0, h 2 (o1)=7, h 3 (o1)=5 BF= h 1 (o2)=0, h 2 (o2)=7, h 3 (o2)=4 √ ×

5. Approximate Counting Multiset: Multiset: Instead of a query – ‘estimate’ operation “How many times the item was added before?” Things to consider: Things to consider: – False positives – No false negatives we only get over approximation! – A tradeoff between space and accuracy Typically solved with Bloom filter extensions – like Spectral Bloom filter, Count Min Sketch, Multi Stage Filters and many more.

6. Counting with Bloom Filter A vector of counters (instead of bits) A counting Bloom filter supports the operations: – Increment Increment by 1 all entries that correspond to the results of the k hash functions – Decrement Decrement by 1 all entries that correspond to the results of the k hash functions – Estimate (instead of get) Return the minimal value of all corresponding entries m=11 368 k=3, h 1 (o1)=0, h 2 (o1)=7, h 3 (o1)=5 CBF= Estimate(o1)=4 4 9 7

7. Some Applications Google Chrome Google BigTable Apache Hadoop Facebook’s) Apache) Casandra Venti (archive system) Cache Admission Policy – And many more…

8. Approximate Set with Hash Tables 1712 We use a bitwise array The function: P assigns each item a place in the array. The function: F assigns each item a fingerprint. – The add operation writes f(o), in the location p(o). – The query operation compares p(o) with the content of f(o) For example: Only works when there are no collisions… × p(o1)=3, h (o1)=7 p(o2)=5, h (o2)=12 √

9. Chain based hash table ? A single pointer is 64 bits – so most of the space is simply for pointers! Array ? Linear Probing? Can we do anything that is more space efficient than an array? Handling Collisions

10. BucketChainTag (fingerprint) Inferred from place in tableOnly tag bits are stored. TinyTable Overview

11. Chain Index Chain Index (1 bit per chain)Array A Is Last Is Last? ( 1 bit per item) BC Chain 2 is not empty Second item is last in chain First Item is not last Chain 7 is empty Encoding of a Single Bucket

12. Chain Index Array ( 6 items of fixed size) Is Last? Logical View Add B to chain 5: A A B B

13. Chain Index Array ( 6 items of fixed size) Is Last? Logical View Add C to chain 0: A A B B A BC C

14. Chain Index Array ( 6 items of fixed size) Is Last? Logical View Add D to chain 2: A B B A BC C BD D A

15. Handling Overflows “When a bucket overflows … ‘steal’ space from a neighboring bucket. “ Bucket 2 Items: 2 Capacity: 5 Bucket 1 Items: 4 Capacity: 5 Bucket 0 Items: 4 Capacity: 5 Bucket 0 Items: 5 Capacity: 5 Bucket 0 Items: 6 Capacity: 6 Bucket 1 Items: 4 Capacity: 4 Bucket 0 Items: 7 Capacity: 7 Bucket 1 Items: 4 Capacity: 4 Bucket 2 Items: 2 Capacity: 4

16. Performance Tradeoff

17. The Tradeoff: Analysis and Empirical

18. Time for 1 million operations (seconds) Alpha = 1.2 Alpha = 1.1 Alpha = 1.025 Query is over 10 times faster than in Bloom filter, regardless of alpha. Update Speed depends on alpha – and can be similar to that of Bloom filters.

19. TinyTable and Counting Bloom Filters This is the original (plain) Bloom filter – without removals. Relative Space Requirement of TinyTable 1.1 Compared to State of the art CBF.

20. TinyTable vs Table Based CBF Alpha = 1.2 Alpha = 1.1

21. Approximate Counting How to represent counters? Consider a single logical chain: We add a single bit per item to indicate whether this item is a fingerprint or a counter part. A This item is a key This item is a counter part. A has a large counter (2 counter parts). B Another key B’s counter Table as a black box: Items can either be counters or keys. Counters are associated with keys to the left.

22. Summery: TinyTable Query is always very fast Query is always very fast. It is based on efficient bitwise operations and very memory local. Update time depends on memory density Update time depends on memory density, denser table = slower update. additions, removals and counting. Full support of additions, removals and counting. smaller than Bloom filters Many attractive configurations! Can be made smaller than Bloom filters with reasonable (better) performance.

23. TinyTable (Alpha =1.2) vs Approximate Counting

24. TinyLFU: Admission policy (PDP 2014) It is not always beneficial to add a new item at the expense of cache victim. Frequency Rank Long Heavy Tail For example~(50% of the weight) A small number of very popular items For example~(50% of the weight)

25. Cache Victim Winner Eviction and Admission Policies Eviction Policy Admission Policy New Item One of you guys should leave… is the new item any better than the victim? What is the common Answer? TinyLFU: Admission policy (PDP 2014)

26. TinyLFU: (PDP 2014) TinyTable Use a sample of recent events to manage cache. (alternative implementation)

27. TinyLFU: Admission policy results 1.Low metadata overhead less than 8 bytes per cache line. 2.Higher cache hit rate 3.Faster cache operation (Query is faster than update)

28. TinyTable is soon to be released as an open source project. TinyLFU was released as the Shades open source project. I believe there are many other applications! Thank you for your Time! (and use my hash table – it is awesome)