CS599: Grid Computing Policy Driven Heterogeneous Resource Co-Allocation with Gangmatching

1
CS599 Grid ComputingPolicy Driven Heterogeneous
Resource Co-Allocation with Gangmatching

• Li Zhou
• zhou_at_isi.edu
• 11/7/2003

2
Guide

• From Matchmaking to Gangmatching
• Why?
• How?
• The Gangmatching Model
• ClassAds
• Usage of Gangmatching (2 Examples)
• Implementation of Matching Algorithm
• Protocol

3
Matchmaking

• Resource allocation or scheduling
• Distributively administrated customer (requester)
  and resource (provider)
• Distributed ownership introduce policy
  heterogenity
• Dynamically define their own data model (use
  application-specific metrics)
• Use Classified Advertisements (ClassAds)

4
Class Advertisements

• Semi-structured data model
• Fold the query language into data model
• Constraint Attribute
• Boolean expression (true/false/undefined)
• The customer / resource pair matches only if the
  expression is evaluated as TRUE
• Rank Attribute
• numeric expression (numval/undefined)
• Choose the matched pair with highest rank value

5
Example

• //Resources ClassAds
• type Machine KFlops 1.8E4 Memory 64
  Arch Intel
• ResearchGroup Friends Untrusted
  
• Rank member (other.Owner, ResearchGroup)10
  member(other.Owner, Friends)
• Constraint !member(other.Owner, Untrusted)
  other.typeJob
• //Requesters ClassAds
• typeJob Ownerzhou
• Rank other.KFlops / 1E4 other.Memory / 32
• Constraint other.TypeMachine
  other.ArchIntel other.KFlopsgt1E4
  other.Memorygtself.Memory

6
Matchmaking ? Gangmatching

• Drawbacks of Matchmaking
• Limited by its purely bilateral formalism
• Can only match a single customer with a single
  resource
• Improvement Gangmatching
• A formalism that extends Matchmaking from
  bilateral model to a multilateral model

7
ClassAds for Gangmatching

• Ports Attribute
• Ports ClassAds
• Define the number and characteristics of matching
  ads required to be satisfied
• Bilateral matching between the ports of ClassAds
• Matchmaking Matching on whole ClassAds
• Label Attribute
• Define the name of candidate that bound to a
  certain port
• Matchmaking use fixed name other

8
Continue

• //Gangmatching ClassAds Definition
• Ports LabelA //classads-a
• Rankrankval-A(A)
• Constraintconsval-A(A)
• LabelB //classads-b
• Rankrankval-B(A,B)
• Constraintconsval-B(A,B)
• LabelN //classads-n
• Rankrankval-N(A,B,,N)
• Constraintconsval-N(A,B,,N)
• //Matchmaking ClassAds Definition
• Rank rankval(other)
• Constraint rankval(other)

9
Usages of Gangmatching

• Condor
• User need licenses for various software package
• Allocate both a machine and a license to a job

Job
A
B
C
D
Machine
License
gcc
z
javac
x
y
10
Corresponding Job ClassAds

• //Jobs ClassAds
• type Job Cmd run_sim
• Ports //request a workstation
• Label cpu
• Rank cpu.KFlops/1E4cpu.memory/32
• Constraint cpu.ArchIntel ,
• //request a license
• Label license
• Host cpu.Name
• Rank license.priority
• Constraint license.typeLicence
  license.appCmd

11
Sophisticated Co-Allocation

• Weather App need large storage to cache
  temporary result
• Constraints imposed on both CPU Storage
• CPU Storage should be on the same host

Job
A
B
C
D
CPU
Storage
a1
a2
b1
c1
d1
d2
12
Corresponding Job ClassAds

• //Jobs ClassAds
• type Job
• Ports //request a workstation
• Label cpu
• Rank cpu.KFlops/1E4cpu.memory/32
• Constraint cpu.ArchIntel ,
• //request a license
• Label storage
• Rank storage.rspeedstorage.wspeed
• Constraint storage.freespacegt1G
  storage.hostcpu.host

13
Implementation

• Naïve Algorithm
• Start from root ClassAds
• Recursive backtracking search
• Until all ports in the gang are bound (and all
  constraints are satisfied)
• Optimization
• Gangmatching problem similar to Join in DB
• ClassAds indexing scheme (attrib. cons.)
• Reordering joins
• (Job Machine) License
• (Job License) Machine
• Adaptive algorithms (choose dynamically)

14
Match Algorithm
Job
CPU
A
D
B
C
Construe Rank9.25 (best)
Consfalse
Construe Rank8.14 (2nd best)
Consfalse
Storage
a1
a2
d1
d2
Construe Rank1000
Construe Rank2000
Consfalse
Consfalse
15
Action Involved in Protocol

• Gangmatching Identical to Matchmakings

Match Algorithm(2)
Gangmatching
Advertisement(1)
Advertisement(1)
Match Notif(3)
Match Notif(3)
Requestor
Providers
Claiming(4)
16
Summary Requirements

• Need to marshal an arbitrary number of candidates
• Ability to relate the properties of multiple
  candidates through arbitrary constraints defined
  on candidate individuals and groups
• Amenable to efficient implementation

17
Summary Characteristics

• Marshal candidate group
• Specific kind of candidate and their
  inter-dependencies are defined only by candidates
  itself
• Matchmaking
• Single implicit bilateral matches
• Gangmatching
• Explicit list of required bilateral matches
• Allow ClassAds to access information from other
  bilateral match localities

18
References

• Rajesh Raman, Miron Livny, Marvin Solomon, Policy
  Driven Heterogeneous Resource Co-Allocation with
  Gangmatching, HPDC 2003 June 22-24
• Rajesh Raman, Miron Livny, Marvin Solomon,
  Matchmaking Distributed Resource Management for
  High Throughput Computing, HPDC 1998
• Classified Advertisements
• http//www.cs.wisc.edu/condor/classad/