sprint: speculative prefetching of remote data

Sprint: Speculative Prefetching of Remote Data

Arun RamanPrinceton University

Greta YorshARM, UK

Martin VechevIBM Research&ETH Zurich

Eran YahavTechnion, Israel

Acknowledgments: Nick Mitchell and Mark Wegman

IBM Research

Client Datasource

RemoteProcessing

(2 sec)

LocalProcessing

(2 sec)

NetworkLatency(16 sec)

Prefetching,Caching

Async,Batching

Query Planning,Cache Opti.

Ted

Chandra

Walter

Heywood

Frank

Rama

Ralph

Dimitri

David

IBM Yellow Pages Application

Remote Access Latency

Compiler

Execution engine

Expose Parallelism across remote accesses

Optimize remote accesses

Sprint (Our Technique)

Node build(String email) { Employee emp = getEmployee(email); if (!emp) return NULL; Node root = new Node(emp); numNodes++; for(reportee_email: emp.getReportees()){ Node child = build(reportee_email);

if (child) { root.addToList(child); child.setParent(root);}

} return root;}

Remote DependencyLocal Dependency

Remote AccessIBM Yellow Pages Application

Ted

Chandra

Walter

Heywood

Frank

Rama

Ralph

Dimitri

David

Ted

Chandra

Walter

Heywood

Frank

Rama

Ralph

Dimitri

David

Program Remote Data- source

input

output

input

Optimist(prefetcher program)

Pessimist(original program)

Sprint execution

engine

cache

output

Remote Data- source

Optimist(prefetcher program)

Pessimist(original program)

• Parallelization•Memory Protection•Output Protection

• Initiating the Optimist• Deadlock Avoidance

Compiler Transformations

Compiler

Execution engine

Expose Parallelism across remote accesses

Optimize remote accesses

Sprint (Our Technique)

Node build(String email) { Employee emp = getEmployee(email); if (!emp) return NULL; Node root = new Node(emp); numNodes++; for(reportee_email: emp.getReportees()){ Node child = build(reportee_email);

if (child) { root.addToList(child); child.setParent(root);}

} return root;}

build(K) { V = get(K) for(k in V.keys) build(k)}

IBM Yellow Pages Application

build(K) { V = get(K) for(k in V.keys) build(k)}

Core 0 Core 1 Core 2

Value StKey

St (State): Absent, Present, or Issued

AAAAAAA

Sprint Cache

Pessimist Optimist

build(K) { V = get(K) for(k in V.keys) build(k)}

Core 0 Core 1 Core 2launch

get(K0)get(K0)

build(K1)get(K1)get(K1)

build(K1)

build(K4)get(K4)

build(K3)get(K3)

build(K3)get(K3)

Pessimist Optimist

Value StKey

St (State): Absent, Present, or Issued

AAAAAAA

K0 I

build(K0) build(K0)

get(K2)build(K2)

get(K2)build(K2)

PV0K1K2

II

V1V2

PP

t0

t1

t2K3K4

II

WAIT

WAIT

HIT!

Sprint Cache

HIT!t3

V3V4

PP

Client Datasource

LocalProcessing

(2 sec)

NetworkLatency(16 sec)

RemoteProcessing

(2 sec)

(2 sec)

Original Execution

(2 sec) (3 sec)Sprint-ed Execution

In the paper: Batching optimization Task prioritization optimization Data access processing algorithm Data consistency with remote updates Correctness proof

• IBM’s Yellow Pages Web Service• Publications Database (DB2)• Facebook Web Service

Datasources• Management Hierarchy• Employee Search• Citation Count• Bibliography Agg.• Friend Connectivity

Clients

QUESTIONS ?

1. IBM’s Yellow Pages Web Service2. Publications Database (DB2)3. Facebook Web Service

Datasources1. Management Hierarchy2. Employee Search3. Citation Count4. Bibliography Agg.5. Friend Connectivity

Clients

Cache Statistics (for Sprint with all optimizations turned on)

Client Accesses Hits Waits Misses Miss % CachedMH 766 747 10 9 1.12% 757ES 293 197 48 48 16.38% 714CC 502 202 168 132 26.29% 370BA 1268 949 178 141 11.12% 1127FC 401 394 0 7 1.75% 394

P1() { x=read(M,a); y=read(M,b); assert (y > x);} P2() { atomic{ write(M,a,2); write(M,b,3) } }

S = {(⟨a, 1⟩, ⟨b, 2⟩), (⟨a, 1⟩, ⟨b, 3⟩), (⟨a, 2⟩, ⟨b, 3⟩)}

read(b) // by Optimist of P1write(a,2),write(b,3) // by P2 read(a) // by Optimist of P1 read(a),read(b) // by Pessimist of P1

S′ = (⟨a, 2⟩, ⟨b, 2⟩)