http://www.grid-support.ac.uk
http://www.ngs.ac.uk
NGS computation services:
API's, concurrent and parallel
jobs
Mike Mineter
mjm@nesc.ac.uk
http://www.nesc.ac.uk/
http://www.pparc.ac.uk/
http://www.eu-egee.org/
Overview
3 distinct techniques will be in the practical:
• API’s to the low-level tools
• Using multiple processors
– Concurrent processing - for many jobs
– Parallel processing – to speed execution of one job
3
Overview
• API’s to the low-level tools
4
Job submission so far
Browser
Portal
Command-line
interfaces
GLOBUS,
GLOBUS,
etc.
etc.
User’s Interface to the grid
5
Application-specific tools
Browser
Browser
Portal
Application Specific
Portal
and / or
Higher generic tools
Command-line
interfaces
GLOBUS, etc.
User’s Interface to the grid
6
Application-specific tools
API’s:
Browser
•Java
•C
Portal
•…
Application Specific
and / or
Higher generic tools
Command-line
interfaces
GLOBUS, etc.
User’s Interface to the grid
7
Available API’s
• “Community Grid” CoG http://www.cogkit.org/
– Java, Python, Matlab
– (very limited functionality on Windows – no GSI)
• C http://www.globus.org/developer/apireference.html
– Example in the practical
8
Using multiple processors
9
High performance
computing
• If run-times are too long, you can
– Optimise, re-code
– Use HPC
• Parallel processing: multiple processors cooperate on one
task – you run one executable; its distributed for you
• Concurrent processing: multiple processors, multiple tasks –
multiple executables distributed and run by a script you write
• High-throughput computing: multiple processors, multiple
tasks using spare CPU cycles - CPU years in elapsed days.
E.g. sensitivity analyses in modelling. Condor, for example.
10
High performance
computing
NGS core nodes open these routes
to you – but you have to do a bit of
work! (Grid is not magic!...)
• HPC
– Parallel processing: on NGS, you run one globus-jobsubmit command – but need to code and build
program so it is parallelised
– Concurrent processing: on NGS, multiple executables
run from a script on the UI
– High-throughput computing: Many independent runs
using spare CPU cycles - CPU years in days.
Watch for addition of
Condor pools on the NGS!!!!
11
Concurrent processing
UI
Globus_job_submit
Internet
Head processors of
clusters
via PBS
Worker processors of
clusters
Processes run independently
12
Concurrent processing
• An approach:
– Script 1: to submit multiple jobs and gather their URIs
from globus-job-submit
– Script 2: to test for completion and gather results
– In the practical, the jobs are short, so these scripts
are integrated
13
Parallel Processing
• In early 1990’s parallel processing had some similarities
to current state of grid computing – if on a smaller scale
–
–
–
–
–
National resources begun to be available
Standards were needed
Lot of hype
Early adopters in physics
Inadequate integration with the desktop
• And from the jungle of conflicting approaches emerged…
• MPI: Message Passing Interface
http://www-unix.mcs.anl.gov/mpi/mpich/
14
Parallel processing
UI
Globus_job_submit
Internet
Head processors of
clusters
Worker processors of
clusters
15
Parallel processing
UI
Internet
Head processors of
clusters
MPI
Worker processors of
clusters
16
Parallel processing
UI
Internet
Head processors of
clusters
MPI
Worker processors of
clusters
Processes communicate, synchronise
(the simpler the better!)
17
MPI concepts
• Each processor is assigned an ID [0,P-1] for P processors
– Used to send/recv messages and to determine processing to be done
• Usually code so that processor 0 has coordinating role:
– Reading input files, Broadcasting data to all other processors
– Receiving results
• Point-to-point Messages are used to exchange data
– Either blocking (wait until both processors are ready to send/receive)
– Or non-blocking: start the send/recv, continue, check for send/recv
completion
• Collective functions
–
–
–
–
Broadcast: sends data to all processors
Scatter: sends array elements to different processors
Gather: collects different array elements from different processors
Reduce: collects data and performs operation as dta are received
• Can group processors for specific purposes (e.g. functional
parallelism)
18
MPI notes
• How does the task split into sub-tasks?
– By functions that can run in parallel??!
– By sending different subsets of data to different processes?
More usual ! Overheads of scatter and gather
• Need to design and code carefully: be alert to
– sequential parts of your program (if half your runtime is
sequential, speedup will never be more than 2)
– how load can be balanced (64 processes with 65 tasks will
achieve no speedup over 33 processes)
– Deadlock!
• MPI functions are usually invoked from C, Fortran
programs, but also Java
• Several example patterns are given in the practical.
Many MPI tutorials are on the Web!
19
Summary
• Can build user applications and higher level
tools on GT2 via APIs
• NGS core nodes support
– concurrent processing (if you do a bit of scripting /
coding)
– MPI for parallelism (if you parallelise your code)
• Remember
– the CSAR, HPCx services!
– NGS is batch oriented at present.
20
Practical
• http://homepages.nesc.ac.uk/~gcw/NGS/GRAM
_II.html
• (That II is letters not numbers on the end!)
21