problem solving with netsolve
DESCRIPTION
Problem Solving with NetSolve. Michelle Miller, Keith Moore, Susan Blackford, NetSolve group, Innovative Computing Lab, University of Tennessee. [email protected] [email protected] [email protected]. www.cs.utk.edu/netsolve [email protected]. NetSolve. Problem Statement - PowerPoint PPT PresentationTRANSCRIPT
Problem Solving with NetSolve
[email protected]@[email protected]
www.cs.utk.edu/[email protected]
Michelle Miller,Keith Moore,Susan Blackford,NetSolve group, Innovative Computing Lab,University of Tennessee
NetSolve
• Problem Statement– Software libraries could be easier to install
and use• Locate library• Configure and install library on local machine
– Need access to bigger/different machines
NetSolve
• Solution– Simple, consistent interface to numeric
packages– Ease of use – locate/use already
configured/installed/running solvers through simple procedure call
– Resource sharing (hardware & software) • Greater access to machine resources• New software packages made available simply
NetSolve ArchitectureServer 1blas, petscAgent
Client Proxy
Server 2lapack, mcell
Server 3blas, itpack
Server 4superLU
Server 5aztec, MA28
MatlabClient
NetSolve ArchitectureServer 1blas, petscAgent
Client Proxy
Server 2lapack, mcell
Server 3blas, itpack
Server 4superLU
Server 5aztec, MA28
MatlabClient
netsolve(‘problemX’, A, rhs)
NetSolve ArchitectureServer 1blas, petscAgent
Client Proxy
Server 2lapack, mcell
Server 3blas, itpack
Server 4superLU
Server 5aztec, MA28
MatlabClient
netsolve(‘problemX’, A, rhs)
Servers?
NetSolve ArchitectureServer 1blas, petscAgent
Client Proxy
Server 2lapack, mcell
Server 3blas, itpack
Server 4superLU
Server 5aztec, MA28
MatlabClient
netsolve(‘problemX’, A, rhs)
Servers?
workload
Server1,Server3
NetSolve ArchitectureServer 1blas, petscAgent
Client Proxy
Server 2lapack, mcell
Server 3blas, itpack
Server 4superLU
Server 5aztec, MA28
MatlabClient
problemX, A, rhs
NetSolve ArchitectureServer 1blas, petscAgent
Client Proxy
Server 2lapack, mcell
Server 3blas, itpack
Server 4superLU
Server 5aztec, MA28
MatlabClient
result
Parallelism in NetSolve
• Task Farming– Single request issued that specifies data
partitioning– Data parallel, SPMD support
Task Farming Interface
/*** BEFORE ***/preamble and initializations;
status1 = netslnb(‘iqsort’, size1, array1, sorted1);status2 = netslnb(‘iqsort’, size2, array2, sorted2);
.
.
.status20 = netslnb(‘iqsort’, size20, array20, sorted20);
program continues;
/*** AFTER ***/preamble and initializations;
status_array = netsl_farm(‘iqsort’, “i=0,19”, netsl_int_array(size_array, “$i”),
netsl_ptr_array(input_array, “$i”), netsl_ptr_array(sorted_array, “$i”));
program continues;
Request Sequencing
• Sequence of computations• Data dependency analysis to reduce
extra data transfers in sequence steps• Transmit superset of all input/output
parameters and make persistent near server(s) for duration of sequence execution.
netsl(“command1”, A, B, C);netsl(“command2”, A, C, D);netsl(“command3”, D, E, F);
Client Server
command1(A, B)
result C
Client Server
command2(A, C)
result D
Client Server
command3(D, E)
result F
netsl_begin_sequence( );netsl(“command1”, A, B, C);netsl(“command2”, A, C, D);netsl(“command3”, D, E, F);netsl_end_sequence(C, D);
Client Server
sequence(A, B, E)
Server
Client Serverresult F
input A,intermediate output C
intermediate output D,input E
Data Persistence
NetSolve Applications
• MCell (Salk Institute)– Monte Carlo simulator of cellular
microphysiology – synaptic transmission– Large numbers of same computation with
different parameters (diffusion and chemical reaction calculations)
– Task farming used for parallel runs
NetSolve and Metacomputing Backends
NetSolveClient
NetSolve Servers
NetSolve Agent
NetSolve Services
MDS
GASSGRAM
Client-ProxyInterface
NetSolveProxy
Client-ProxyInterface
GlobusProxy
Client-ProxyInterface
NinfProxy
LegionProxy
NinfServices Legion
Services
Client-Proxy Interface
NetSolve Authenticationwith Kerberos
• Kerberos used to maintain Access Control Lists and manage access to computational resources.
• NetSolve properly handles authorized and non-authorized components together in the same system.
NetSolve Authenticationwith Kerberos
NetSolve client
NetSolve agent
NetSolve servers
KerberosKDC
Typical NetSolve Transaction
Kerberized Interaction
NetSolve Authentication with Kerberos
NetSolve client
NetSolve agent
NetSolve servers
KerberosKDC
Servers registertheir presencewith the agent
and KDC
Client issues problem request;Agent responds with list of servers
Client sends work request to server; server replies requesting
authentication credentials
Client requests ticket from KDC
Client sends ticket and input to server; server authenticates and
returns the solution set
NWS Integration
NetSolveServer CPU
sensor
Host Machine
NetSolveServer CPU
sensor
Host MachineHost Machine
NetSolveAgent
NWSForecaster
NWSMemory
Sensors report to NWS memory.
NWS Integration
NetSolveServer CPU
sensor
Host Machine
NetSolveServer CPU
sensor
Host MachineHost Machine
NetSolveAgent
NWSForecaster
NWSMemory
NWS Integration
NetSolveServer CPU
sensor
Host Machine
NetSolveServer CPU
sensor
Host MachineHost Machine
NetSolveAgent
NWSForecaster
NWSMemory
Agent probes NWS Forecaster
NWS Integration
NetSolveServer CPU
sensor
Host Machine
NetSolveServer CPU
sensor
Host MachineHost Machine
NetSolveAgent
NWSForecaster
NWSMemory
Forecaster probes memory.
NWS Integration
NetSolveServer CPU
sensor
Host Machine
NetSolveServer CPU
sensor
Host MachineHost Machine
NetSolveAgent
NWSForecaster
NWSMemory
Forecaster makes forecast
Agent chooses server
NWS Integration
NetSolveServer CPU
sensor
Host Machine
NetSolveServer CPU
sensor
Host MachineHost Machine
NetSolveAgent
NWSForecaster
NWSMemory
Newly enabled libraries
PETSc
SuperLU
SPOOLES
MA28
Sparse Matrices/Solvers• Support for compressed row/column sparse
matrix storage -- significantly reduces network data transmission.
• Iterative and direct solvers: PETSc, Aztec, SuperLU, Ma28, …
• All available solver packages will be made available from UTK NetSolve servers and others.
Matlab interface
• Calls to PETSc, Aztec– [x, its] = netsolve(‘iterative_solve_parallel’,‘PETSC’,A,b,1.e-6,500);– [x, its] = netsolve(‘iterative_solve_parallel’, ‘AZTEC’,A,b,1.e-6,500);
• Similar for SuperLU, MA28– [x] = netsolve(‘direct_solve_serial’,’SUPERLU’,A,b,0.3,1);– [x] = netsolve(‘direct_solve_serial’, ‘MA28’,A,b,0.3,1);
• Calls to LAPACK, ScaLAPACK– [lu, p, x, info] = netsolve(‘dgesv’,A,b);– [lu,p,x,info] = netsolve(‘pdgesv’,A,b);
‘LinearSolve’ interface
• Uncertain which library to choose?• ‘LinearSolve’ interface chooses the library
and appropriate routine for the user– [x] = netsolve(‘LinearSolve’,A,b);
Heuristics
• Interface analyzes the matrix A– matrix shape (square, rectangular)?
• If rectangular, choose linear least squares solver from LAPACK
– matrix element density• If square, is the matrix sparse or dense?• Manually check the percentage of nonzeros and
transform to sparse matrix format• If dense, is it symmetric?
Heuristics cont’d
• If sparse, and dense band, use LAPACK• If sparse, and if there is a block or diagonal
structure (Aztec), can yield higher performance (level 3 BLAS)
• If sparse, direct or iterative solver?– Size of the matrix (large matrix, fill-in for a
direct method can be larger than iterative)
Heuristics cont’d
– Numerical properties (direct solvers can handle more complicated matrices than iterative methods)
• How to estimate fill-in and gauge numerical properties of A?
• Future Work: ‘Eigensolve” interface
Interfacing to Parallel Libraries
• Improved task scheduling and load balance– NWS memory sensor, latency and bandwidth
sensors– provide matrix distributions for the user– heuristics for the best choice of the # of
processors, amount of matrix per processor, process grid dimension ...
Get NetSolve1.3 Now!
• Release date -- April 2000!• UNIX client/agent/server source code.• UNIX client binaries available.• Win32 dlls for C/Matlab/Mathematica
clients.
www.cs.utk.edu/netsolve