Robust Grid Computing Using PeertoPeer Services

1
Robust Grid Computing Using Peer-to-Peer
Services

• Alan Sussman
• Department of Computer Science

2
Desktop Grids and P2P Systems
3
P2P Desktop Grid System

• Executing applications on a (widely) distributed
  set of resources
• Decentralized, Robust, Highly available and
  Scalable
• Goal is to enable resource sharing
• create ad-hoc grids within (research) communities
• Applications
• Large computational requirements and relatively
  low I/O requirements
• Physical simulations and data analysis related to
  astronomy, physics, other scientific disciplines
  - parameter sweeps
• binary asteroid formation (co-I Richardson)
• comet composition (Deep Impact, co-I Wellnitz)
• Monte Carlo simulations
• Bioinformatics applications

4
Hard Problems / Issues

• Job Submission
• Submit a job into the decentralized P2P system
• Matchmaking
• Find a resource that can meet the minimum
  resource requirements of a job
• Load balance
• Distribute the load (jobs) across the nodes in
  the system
• And no node has too much overhead (messages)
• Resilience to failure
• Overall system must not lose jobs from failures,
  and allow nodes to join and leave (or fail)
  dynamically

5
System Architecture
Job J
Peer-to-Peer Network (DHT)
Clients
Assign GUID to Job J
6
Goals of Matchmaking Algorithm

• Expressiveness
• Allow users to specify any type of resource
  requirements
• Load balance
• Distribute load across multiple capable
  candidates
• Parsimony
• Resources should not be wasted
• Completeness
• A valid assignment of a job to a node must be
  found if such an assignment exists
• Low overhead
• Routing must not add significant overhead

7
Basic Assumptions

• Underlying Distributed Hash Table (DHT)
• Object location and routing in P2P network
• Reformulate the problem of matchmaking to one of
  routing
• Job in the system
• Data and associated profile (requirements)
• All jobs are independent
• Optimization criterion
• Minimize time to complete all jobs (combination
  of throughput and response time)
• Queuing time used as proxy for response time

8
Modified Content-Addressable Network

• Basic CAN
• Logical d-dimensional space
• zone, neighbors, greedy forwarding
• Formulate the matchmaking problem as a routing
  problem in CAN space
• Each resource type is a CAN dimension
• Map nodes and jobs into the CAN space
• Resource capabilities and requirements,
  respectively
• Search for a node whose coordinates in all
  dimensions meet or exceed the jobs requirements

9
CAN-based Algorithm

• Employ dynamic aggregated resource information
  for load balancing

Memory Dimension

• Aggregate Resource Information

• Choose the least loaded direction

• Push a job into under-loaded region

• Stop pushing

Node A

• Choose the best run node

Job
Routing
CPU Dimension
10
Different Resource Types

• Categorical Resources
• Architecture, Operating system type (version),
  etc.
• Exact match
• Continuous Resources
• CPU speed, Memory amount, Disk space, etc.
• Minimum match
• Until recently, our work considered only
  continuous resource types

11
Integrating Categorical Resources

• How to integrate all types of resources in CAN?
• Could add new CAN dimensions for categorical
  resource types
• Could create a distinct CAN space for each
  combination of categorical resource types
• Our Approach
• Virtual Peers 1-dimensional transformation

12
Virtual Peer1-D Transformation

• 1-dimensional Transformation
• Transform all categorical resource dimensions
  into a single dimension
• To make virtual peer management and failure
  recovery simple, and re-use our existing
  algorithms
• Use Hilbert Space-Filling Curve
• Virtual peers cover the unoccupied parts of the
  CAN space in the categorical dimension
• Effectively divide the CAN space into multiple
  disjoint sub-CANs and connect them

13
Memory
A
G
E
B
H
VBE
VEH
VH
MJ
C
F
I
D
Linux Intel
OSX PPC
Windows Intel
AIX Power
Solaris Sparc
C- Dim
14
Improving Overall Scalability

• Partial Updates
• Limit the size of a single update message
• Probabilistic Heartbeat Messaging
• Limit the number of outgoing update messages
• Specialized Routing along T-dimension
• Balance the load of processing routing messages
  from a sub-CAN to another sub-CAN
• All without affecting correctness of CAN
  algorithms
• Must quantify performance and reliability effects

15
Experimental Results
16
Current Status

• Resource discovery algorithms thoroughly
  simulated and verified
• CAN-based implementation complete and being
  tested
• All CAN services working node join, node
  leave/fail, job assignment, load aggregation
• Basic authentication mechanism in place, based on
  certificates and public-key authentication
• Job management and client interface completed
• Categorical resource types still being
  implemented
• Large-scale deployment and measurement in
  process
• binary asteroid formation application (Richardson)

17
Ongoing Work

• Deploying the prototype system for real
  workloads and real machines
• practical issues, such as firewalls, user
  interface for job submission
• with serious logging for measuring system
  behavior
• Dealing with multicore and multiprocessor
  machines
• Better characterization of real workloads
• via consultation with Astronomy collaborators,
  logging behavior of the peer, and mining of
  Condor system logs
• High quality peer software, with good user tools

18
The Project Team

• Faculty members
• Alan Sussman, Pete Keleher, Bobby Bhattacharjee,
  Derek Richardson (Astronomy), Dennis Wellnitz
  (Astronomy)
• Matchmaking algorithms and simulations
• Jik-Soo Kim (RA)
• Jaehwan Lee (RA)
• Peer implementation and other tools
• Michael Marsh (research scientist), Beomseok Nam
  (postdoc)
• Sukhyun Song, San Ratanasanya (RAs)

19
End of Talk