Transcript Internet Backplane Protocol API and Applications

Logistical Computing and
Internetworking: Middleware for the Use
of Storage in Communication
Micah Beck
Jack Dongarra
Terry Moore
James Plank
University of
Tennessee
Fran Berman
Henri Casanova
Rich Wolski
University of
California, San
Diego
University of
California,
Santa Barbara
3rd International Workshop on Active Middleware Services
1
LoCI Projects
•
•
•
•
Internet Backplane Protocol (Beck, Plank)
Network Weather Service (Wolski)
NetSolve (Dongarra)
Application Level Scheduling (Berman)
LoCI Funded by National Science Foundation
Next Generation Software Program
2
Internet Backplane Protocol (IBP)
• primitive middleware that supports a layer
of network storage
• implemented as a system of buffers exposed
for direct scheduling,
• can be used by advanced applications to
leverage state management for highperformance.
3
Network Weather Service (NWS)
• Monitors and extrapolates network metrics
– network bandwidth and latency
– storage availability
– CPU load
• Prediction is weak reservation
– all reservations will sometimes be broken
– effective for highly aggregated resources
4
NetSolve (NetSolve)
• Provides a programming environment that
facilitates the analysis of program
dependences to understand an application’s
inherent communication requirements.
• A major component of LoCI research is
identify and provide opportunities for
extracting scheduling information from
applications.
5
Application Level Scheduling
(AppLeS)
• Enables the derivation of an efficient schedule that
matches communication requirements.
• Mapping the computation, network and storage
resources of the application to the Grid resources
subject to current and predicted resource
conditions, is a difficult problem.
• AppLeS is the leading instance of a range of
approaches we are exploring under LoCI.
6
An Analogy with Pipelined
Processor Architecture
• The fundamental elements of modern
processor architecture are:
– Buses and functional units which move and
transform data, and
– Memory and cache, registers and pipeline
buffers that store data.
• RISC architecture exposes resources to
scheduling by the compiler
7
Network Computing has
Analogous Components
• In our model of logistical network
computing, the fundamental elements are
– Predictable networking and computation which
move and transform data, and
– Storage that is accessible from the network.
• Logistical Computing exposes resources to
external schedulers (including applications)
8
Logistical Networking:
Exposed Storage Management
• Storage resources available for direct access
at network intermediate nodes.
• Allocation and scheduling of storage
resources are exposed to the network.
• Some implications
– storage resources are shared among operations
– applications, intermediate nodes can schedule
9
IBP Software Structure
• IBP Depots (servers) are daemons that serve
local storage to IBP clients.
• IBP Clients link an IBP client library with a
well-defined API.
• Clients talk to depots using TCP/IP.
• Design is for high-performance/scalability.
10
Logistical Computation
Mechanisms
• The Network Weather Service: Monitoring
Resources for Logistical Scheduling
• Logistical Scheduling and the AppLeS
Project
• Coscheduling of Storage and Computation
in NetSolve
11
NetSolve - The Big Picture
Computational Resources
Clusters
Reply
Choice
MPP
Workstations
MPI, PVM,Condor...
Matlab
Mathematica
C, Fortran
Java, Perl
Java GUI
Agent
Request
Scheduler
Database
Client - RPC like
12
State Management
in NetSolve
• The Problem: NetSolve
calls are functional
• Excessive data transfers
For example:
Client
A,B
Server 1
F
X
Client
X,B
Y
Server 2
G
Client
X = F(A, B);
Y = G(X, B);
13
Dependence
Flow
Caching
Client
A,B
Server 1
A,B
X
Client
B
F
IBP Cache
B
B
X,B
Server 2
Y
G
Server 1
F
X
Server 2
Y
Client
A
Y
G
Client
14
An Experiment Using NetSolve
• NetSolve Client at UC San Diego
• Computational Servers at UT Knoxville
• MA28 solver library used to solve systems
of equations from the Harwell-Boeing
collection of the Matrix Market repository
• Uncached to client-directed caching
15
Preliminary Results
• Unenhanced NetSolve vs.
NetSolve w/IBP caching
16.1 KB
2.68 MB
16
LoCI Software Integration
• IBP Depot (server) available for Unix/Linux
and Win32
• IBP Client Library also available for Java
• NetSolve 1.4 (just released) supports IBP
caching
• Network Weather Service uses IBP
internally for monitor state management
17
Conclusions
• Logistical Computing defines a
comprehensive exposed approach to Grid
computing
• Processing, network, and storage resources
are explicitly scheduled for performance
• Storage resources sharing enables
improvements over stateless computation
based solely on end-to-end communication
18