Optimal Fast Hashing Yossi Kanizo (Technion, Israel) Joint work with Isaac Keslassy (Technion, Israel) and David Hay (Hebrew Univ., Israel)

Hash Tables for Networking Devices  Hash tables and hash-based structures are often used in high-speed devices  Heavy-hitter flow identification  Flow state keeping  Flow counter management  Virus signature scanning  IP address lookup algorithms

Hash tables  In theory, hash tables are particularly suitable: O(1) memory accesses per operation (element insertion/query/deletion) for reasonable load  But in practice, there is a big difference between an average of 1.1 memory accesses per operation, and an average of 4  Why not only 1 memory access?  Collisions

Hash Tables for Networking Devices 123  Collisions are unavoidable  wasted memory accesses  For load≤1, let a and d be the average and worst- case time (number of memory accesses) per element insertion Objective: Minimize a and d Memory

Why We Care  On-chip memory: memory accesses  power consumption  Off-chip memory: memory accesses  lost on/off-chip pin capacity  Datacenters: memory accesses  network & server load  Parallelism does not help reduce these costs  d serial or parallel memory accesses have same cost

Traditional Hash Table Schemes  Example 1: linked lists (chaining) Memory

Traditional Hash Table Schemes  Example 1: linked lists (chaining)  Example 2: linear probing (open addressing)  Problem: the worst-case time cannot be bounded by a constant d Memory

High-Speed Hardware  Enable overflows: if time exceeds d → overflow list  Can be stored in expensive CAM  Otherwise, overflow elements = lost elements  Bucket contains h elements  E.g.: 128-bit memory word  h=4 elements of 32 bits  Assumption: Access cost (read & write word) = 1 cycle Memory h CAM 9

Possible Settings  Static setting - Insertions and queries only  Dynamic setting – Insertions, deletions, and queries.  Generalized setting – Balancing between the buckets’ load.

Problem Formulation Memory h CAM 9 Given average a and worst-case d of memory accesses per operation, Minimize overflow rate  Given average a and worst-case d of memory accesses per operation, Minimize overflow rate 

Example: Power of d-Random Choices  d hash functions: pick least loaded bucket.  Break ties u.a.r. [Azar et al.]  Intuition: can reach low  … but average time a = worst-case time d  wasted memory accesses Memory h CAM

Other Examples  d-left [Vöcking]  Same as d-random, but break ties to the left.  Cuckoo [Pagh et al.]  Whenever collision occurs, moves stored elements to their other choices.  Typically, uses much more than d memory accesses on average.

Outline  Static Case  Overflow Lower Bound  Optimal Schemes: SIMPLE, GREEDY, MHT.  Dynamic Case  Comparison with Static Case.  Overflow Lower Bound  Overflow Fraction Depending on d.

Overflow Lower Bound  Objective: given any online scheme with average a and worst-case d, find lower-bound on overflow . [h=4, load=n/(mh)=0.95, fixed d] No scheme can achieve (capacity region)

Overflow Lower Bound  Result: closed-form lower-bound formula  Given n elements in m buckets of height h:  Valid also for non-uniform hashes  For n=m and h=1, we get simply  Defines a capacity region for high- throughput hashing

Lower-Bound Example [h=4, load=n/(mh)=0.95] For 3% overflow rate, throughput can be at most 1/a = 2/3 of memory rate

Overflow Lower Bound  Example: d-left scheme: low overflow , but high average memory access rate a [h=4, load=n/(mh)=0.95, m=5,000]

The SIMPLE Scheme  SIMPLE scheme: single hash function  Looks like truncated linked list Memory h CAM

Performance of SIMPLE Scheme [h=4, load=0.95, m=5,000] The lower bound can actually be achieved for a=1

The GREEDY Scheme  Using uniform hashes, try to insert each element greedily until either inserted or d Memory h CAM d=2

Performance of GREEDY Scheme [d=4, h=4, load=0.95, m=5,000] The GREEDY scheme is always optimal until a co

Performance of GREEDY Scheme [d=4, h=4, load=0.95, m=5,000] Overflow rate worse than 4-left, but better throughput (1/a)

The MHT Scheme  MHT (Multi-Level Hash Table) [Broder&Karlin]: d successive subtables with their d hash functions Memory h CAM st Subtable2 nd Subtable3 rd Subtable

Performance of MHT Scheme  Optimality of MHT until cut-off point a co (MHT)  Proof that subtable sizes fall geometrically  Confirmed in simulations [d=4, h=4, load=0.95, m=5,000] Overflow rate close to 4-left, with much better throughput (1/a)

Outline  Static Case  Overflow Lower Bound  Optimal Schemes: SIMPLE, GREEDY, MHT.  Dynamic Case  Comparison with Static Case.  Overflow Lower Bound  Overflow Fraction Depending on d.

Dynamic vs. Static  Dynamic hash tables are harder to model than the static ones [Kirsch et al.]  But past studies show same asymptotic behavior with infinite buckets (insertions only vs. alternations)  traditional hashing using linked lists – maximum bucket size of approx. log n / log log n [ Gonnet, 1981]  d-random, d-left schemes – maximum bucket size of log log n / log 2 + O(1) [ Azar et al.,1994; Vöcking, 1999]  As a designer, using the static model seems natural.  Even if real-life devices have finite buckets

Degradation with Finite Buckets  Finite buckets are used.  Surprising result: degradation in performance FiniteInfinite H(1) = 3H(2) = 3 Remove 1 Element “2” is lost although its corresponding bucket is empty

Comparing Static and Dynamic  Static setting: insertions only  n = number of elements  m = number of buckets  Dynamic setting: alternations between element insertions and deletions of randomly chosen elements.  fixed load of c = n / (mh)  Fair comparison  Given an average number of memory accesses a, minimize overflow fraction .

Overflow Lower Bound  Overflow lower bound of where r = ach.  Also holds for non-uniformly distributed hash functions (under some constraints).  The lower bound is tight (Simple, Greedy)

Numerical Example  For h=1 and c=1 (100% load) we get a lower bound of 1/(1+a).  To get an overflow fraction of 1%, one needs at least 99 memory accesses per element.  Infeasible for high-speed networking devices  Compared to a tight upper bound of e -a in the static case. [Kanizo et al., INFOCOM 2009]  need ~4.6 memory accesses.

Outline  Static Case  Overflow Lower Bound  Optimal Schemes: SIMPLE, GREEDY, MHT.  Dynamic Case  Comparison with Static Case.  Overflow Lower Bound  Overflow Fraction Depending on d.

Overflow Fraction Depending on d  So far, we relaxed the constraint on d.  We considered n elements with an average of a memory accesses, as n  a distinct elements.  To take into account d, we must consider each element along with its own hash values.

Graph Theory Approach  Consider a bipartite graph.  Left vertices = Elements  Right vertices = Buckets (assume h=1).  Edge = The bucket is one of the element’s d choices

Graph Theory Approach  We get a random bipartite graph where each left vertex has degree d.  Expected maximum size matching = Expected number of elements that can be inserted to the table, that is, a lower bound.  We derived an explicit expression for d=2.  Upper bound can be achieved by Cuckoo hashing (equivalent to finding maximum size matching).

Summary  We found lower and upper bounds on the achievable overflow fraction both for the static and dynamic cases.  Static models are not necessarily exact with dynamic hash tables.  Improved lower bound for d=2 and a characterization of the performance of Cuckoo hashing.

Thank you.