Distributed and Non-Distributed Computational Models Ami Paz IRIF β CNRS and Paris Diderot University
Message Passing Models Local 1. Congest 2. Clique 3.
Message Passing Models ο½ A graph π» = π, πΉ representing the network β s topology ο½ π unbounded processors, located on the nodes ο½ Communicating on the edges ο½ Synchronous network ο½ Compute / verify graph parameters 3
The Local Model ο½ Unbounded messages ο½ Solving local tasks: ο½ Coloring ο½ MST ο½ MIS 2-hop environment ο½ Anything solvable in Ξ πΈ rounds * 1-hop environment 4
Two Examples ο½ Triangle detection ο½ Easy, in one round ο½ Send all your neighbors your list of neighbors ο½ Computing the diameter πΈ ο½ Takes Ξ(πΈ) rounds 5
Diameter Lower Bound ο½ Computing πΈ takes Ξ© πΈ rounds ο½ Indistinguishability argument πΈ = π/2 πΈ = π β 1 6
Diameter Lower Bound ο½ Computing πΈ takes Ξ© πΈ rounds ο½ Indistinguishability argument View after π/2 β 1 rounds πΈ = π/2 View after π/2 β 1 rounds πΈ = π β 1 Cannot distinguish 7
The Congest Model ο½ Bounded message size; typically π = O log π ο½ All Local lower bounds still hold ο½ Some Local algorithms still work ο½ But not all! Bottleneck 8
Congest β Typical Lower Bound [HW12] ο½ Communication complexity problem ο½ Inputs encoded by a graph ο½ Split the graph between Alice and Bob ο½ CC lower bounds imply message lower bounds 0 Disjointness on 1 Ξ π 2 bits. 2 Diam 2 or 3 ? 3 β’ Diam 2 β disjoint 4 β’ Diam 3 β not disjoint Ξ© (π) rounds are needed Alice Bob Bottleneck
Congest β Another Lower Bound Alice Ξ©( π/π) lower bound Bottleneck Verification: MST, bipartiteness, cycle, connectivity β¦ Approximation: MST, min cut, shortest s-t path β¦
So Far: ο½ Local model: ο½ Unbounded messages ο½ Everything is solvable in π πΈ rounds ο½ Congest model: ο½ Message = π log π bits ο½ Lower bounds of Ξ© π + πΈ ο½ Tight for many problems ο½ Question: is Ξ© π due to congestion? 11
The Clique Model ο½ All-to-all message passing β a clique network ο½ Diameter of 1 ο½ No distance β only congestion ο½ MST in π(log β π) rounds [GP16] ο½ Fast triangle detection, diameter, APSP, β¦ 12
Clique β Lower Bound? ο½ Diam = 1 ο½ Larger set β more outgoing edges ο½ No nontrivial lower bound is known ο½ Simple counting argument [DKO14] ο½ many functions need π β 5 log π rounds 13
Parallel Systems
Parallel Systems ο½ π synchronous processors, π inputs to each ο½ Connected by a communication graph ο½ Typical graphs: ο½ Clique ο½ Cycle ο½ T orus (Grid) ο½ Known topology, known identities ο½ Bounded message size ο½ Bounded memory ο½ Bounded computational power 15
Parallel vs. Congest ο½ Parallel is more restrictive: ο½ Bounded memory ο½ Bounded computational power ο½ Different focus: ο½ Specific communication graphs ο½ Algebraic questions vs. graph parameters 16
Circuits
Circuits ο½ Algebraic computation model ο½ A computation graph (circuit) composed of: ο½ Inputs, output, and operation gates ο½ Represent many algorithms: ο½ Matrix multiplication, determinant, permanent ο½ Complexity measures: ο½ Depth, number of gates, fan-in, fan-out Λ * Λ Λ Λ + + + 18
Circuits Families ο½ Arithmetic circuits ο½ Boolean circuits ο½ Boolean circuits augmented with: ο½ mod π gates ο½ Threshold gates ο½ β¦ mod 3 Λ Λ Λ Λ 19
Circuits Lower Bounds ο½ What can be computed in constant depth? ο½ Counting argument: ο½ Many functions cannot be computed using Boolean circuits ο½ β¦ or even using augmented circuits ο½ But: ο½ No explicit function is known 20
Circuits β Clique
Clique vs. Circuits ο½ Clique can simulate circuits [DKO14] ο½ Each node simulates a set of gates in a layer ο½ Circuit β s depth = # of rounds Λ Λ mod 3 mod 3 Λ Λ Λ Λ Λ Λ Λ Λ 22
Clique vs. Circuits ο½ Main idea: ο½ Simulate each layer of the circuit in π 1 rounds Λ Λ mod 3 mod 3 π§ Λ Λ Λ Λ Λ Λ Λ Λ Λ π¦ π¦ π§ 23
Clique vs. Circuits ο½ Main idea: ο½ Simulate each layer of the circuit in π 1 rounds Λ Λ mod 3 mod 3 Λ Λ Λ Λ Λ Λ Λ Λ 24
Clique vs. Circuits ο½ Main idea: ο½ Simulate each layer of the circuit in π 1 rounds Λ Λ mod 3 mod 3 Λ Λ Λ Λ Λ Λ Λ Λ 25
Clique vs. Circuits ο½ Main idea: ο½ Simulate each layer of the circuit in π 1 rounds Λ Λ mod 3 mod 3 Λ Λ Λ Λ Λ Λ Λ Λ 26
Clique vs. Circuits ο½ Clique can simulate circuits ο½ Non-constant rounds lower bound for the Clique β Non-constant depth lower bound for circuits ο½ There is also a reduction in the other direction [DKO14] ο½ A circuit can simulate the Clique mod 3 Λ Λ Λ 27
Parallel β Clique
Matrix Multiplication ο½ Base for many algebraic problems ο½ Thoroughly studied in parallel computing ο½ Several algorithms: ο½ different topologies, input / output partitions = β π π π 29
Skip Details Matrix Multiplication ο½ This talk: ο½ The 3D algorithm [ABG+95] ο½ For π Γ π matrices and π processors ο½ Adaptation of parallel algorithm to the Clique [CHK+16] = β π π π 30
Matrix Multiplication ο½ Parallel 3D algorithm β ο½ Clique matrix multiplication in π π 1/3 rounds ο½ Implies triangle detection, πΈ , APSP , β¦ ο½ In similar time [CHK+16] = β π π π 31
Fast Matrix Multiplication ο½ Standard matrix multiplication: ο½ Compute π 2 entries, each need π multiplications otal: Ξ π 3 time ο½ T ο½ There exist faster algorithms: ο½ Strassen π π 2.807 [1969] ο½ Coopersmith-Vinograd π π 2.376 [1990] ο½ β¦ ο½ Le Gall π π 2.373 [2014] ο½ Can be implemented in the Clique ο½ Distributed matrix multiplication in π(π 0.158 ) rounds 32
Some Results & Conclusion
Triangle Detection in the Clique 1. Combinatorial algorithm: 1 3 ο½ Ξ π rounds [DLP12] 2. Reduction from circuits for matrix multiplication: π πβ2 β Ξ π 0.373 rounds, randomized [DKO14] ο½ 3. Using a technique from parallel matrix multiplication: 2 π β Ξ π 0.158 rounds [CHK+16] ο½ O π 1β ο½ 2,3 Imply similar complexities for: ο½ APSP, diameter, girth Sequential matrix multiplication: π π π operations 34
Conclusion ο½ Several models: ο½ Message passing * ο½ Local , Congest and Clique + + + ο½ Parallel systems ο½ Circuits ο½ Arithmetic, Boolean, augmented ο½ Many connections and similarities ο½ Approach different questions ο½ Using different techniques Thank You! 35
Recommend
More recommend