C2115 Practical Introduction to Supercomputing 7th Lesson -1C2115
Practical Introduction to
Supercomputing
Petr Kulhánek, Jakub Štěpán
kulhanek@chemi.muni.cz
National Centre for Biomolecular Research, Faculty of Science
Masaryk University, Kotlářská 2, CZ-61137 Brno
CZ.1.07/2.2.00/15.0233
7th Lesson
C2115 Practical Introduction to Supercomputing 7th Lesson -2-
Contents
 Exercise LV.2
results, discussion
 Computer architecture
limiting factors, application types and relation to limiting factors
 Exercise
network transfer measurements
 Batch systems
definition, overview
C2115 Practical Introduction to Supercomputing 7th Lesson -3Exercise
LV.2
 Results
 Discussion
C2115 Practical Introduction to Supercomputing 7th Lesson -4-
Results
wolf01, 4 CPU, Intel(R) Xeon(R) CPU X3460 @ 2.80GHz, L1: 32kB, L2: 256kB, L3: 8192kB
load_cpu
Number of
concurrent
processes
Real runtime
[s]
Theoretical runtime
[s]
Overhead [%]
1 20.15 20.15
4 30.20 20.15 49.9
8 61.67 40.30 53.0
12 94.12 60.45 55.7
16 126.23 80.60 56.6
20 159.87 100.75 58.7
24 191.64 120.90 58.5
100100 
theory
real
t
t
overhead How many % slower is run to theoretical
ideal.
C2115 Practical Introduction to Supercomputing 7th Lesson -5-
Results
wolf01, 4 CPU, Intel(R) Xeon(R) CPU X3460 @ 2.80GHz, L1: 32kB, L2: 256kB, L3: 8192kB
load_cpu
Number of
concurrent
processes
Real runtime
[s]
Theoretical runtime
[s]
Overhead [%]
1 20.15 20.15
4 30.20 20.15 49.9
8 61.67 40.30 53.0
12 94.12 60.45 55.7
16 126.23 80.60 56.6
20 159.87 100.75 58.7
24 191.64 120.90 58.5
Overheadgrows
100100 
theory
real
t
t
overhead How many % slower is run to theoretical
ideal.
C2115 Practical Introduction to Supercomputing 7th Lesson -6-
Results
wolf01, 4 CPU, Intel(R) Xeon(R) CPU X3460 @ 2.80GHz, L1: 32kB, L2: 256kB, L3: 8192kB
load_cpu
Number of
concurrent
processes
Real runtime
[s]
Theoretical runtime
[s]
Overhead [%]
1 20.15 20.15
4 30.20 20.15 49.9
8 61.67 40.30 53.0
12 94.12 60.45 55.7
16 126.23 80.60 56.6
20 159.87 100.75 58.7
24 191.64 120.90 58.5
Find reason for high
start overheads
100100 
theory
real
t
t
overhead How many % slower is run to theoretical
ideal.
Overheadgrows
C2115 Practical Introduction to Supercomputing 7th Lesson -7Computer
architecture
 Limiting factors
 Application types and their relation to
limiting factors
C2115 Practical Introduction to Supercomputing 7th Lesson -8Architecture,
general view
CPU
North bridge
South
bridge
Memory
SATA controller
Hard drives
Network (ethernet)
C2115 Practical Introduction to Supercomputing 7th Lesson -9Architecture,
limiting factorsfaktory
CPU
North bridge
South
bridge
Memory
Network (ethernet)
SATA controller
Hard drives
cache
Fastest component is CPU
Other components are slower
RAM
~10 GB/s
SATA disc
SATA III: 600 MB/s
Network
10/100/1000 Mb/s
Limiting factors
C2115 Practical Introduction to Supercomputing 7th Lesson -10Architecture,
limiting factors
CPU
North bridge
South
bridge
Memory
Network (ethernet)
SATA controller
Hard drives
cache
Fastest component is CPU
Other components are slower
RAM
~10 GB/s
SATA disc
SATA III: 600 MB/s
Network
10/100/1000 Mb/s
Limiting factors
High latencies
C2115 Practical Introduction to Supercomputing 7th Lesson -11-
1. Use command wget to download install image of Ubuntu Server 12.04.1 LTS
2. State transfer rate for different number of concurrent downloading processes in teams.
What is the limiting factor for the transfer?
Exercise VI.1
$ wget http://www.ubuntu.com/start-download?distro=server&bits=64&release=lts
C2115 Practical Introduction to Supercomputing 7th Lesson -12Batch
systems
 Definition
 Overview
C2115 Practical Introduction to Supercomputing 7th Lesson -13Batch
processing
Batch processing is running of series of programs (so called batches) on computer with no
user interaction. Batches are prepared in advance and submitted for processing without
user interaction. All input data are prepared in advance in files (scripts) or given in
command line arguments. Batch processing is opposite to interactive processing when user
gives input data during actual program run.
Batch processing advantages
 resource sharing among many users and programs
 start of batch run when computer has enough resources (low load)
 remove of user input delays
 maximizing computer usage improves computer investments utilization (expensive
machines)
Source: www.wikipedia.cz, adjusted
C2115 Practical Introduction to Supercomputing 7th Lesson -14Batch
system tools
OpenPBS
http://www.mcs.anl.gov/research/projects/openpbs/
PBSPro
http://www.pbsworks.com
Oracle Grid Engine
http://www.oracle.com/us/products/tools/oracle-grid-engine-075549.html
Open Grid Scheduler
http://gridscheduler.sourceforge.net/
Torque
http://www.adaptivecomputing.com/products/open-source/torque/
C2115 Practical Introduction to Supercomputing 7th Lesson -15Batch
system tools
OpenPBS
http://www.mcs.anl.gov/research/projects/openpbs/
PBSPro
http://www.pbsworks.com
Oracle Grid Engine
http://www.oracle.com/us/products/tools/oracle-grid-engine-075549.html
Open Grid Scheduler
http://gridscheduler.sourceforge.net/
Torque
http://www.adaptivecomputing.com/products/open-source/torque/
open source
C2115 Practical Introduction to Supercomputing 7th Lesson -16Batch
system tools
OpenPBS
http://www.mcs.anl.gov/research/projects/openpbs/
PBSPro
http://www.pbsworks.com
Oracle Grid Engine
http://www.oracle.com/us/products/tools/oracle-grid-engine-075549.html
Open Grid Scheduler
http://gridscheduler.sourceforge.net/
Torque
http://www.adaptivecomputing.com/products/open-source/torque/
open source
Used as batch system in MetaCentrum VO, and on clusters SOKAR and WOLF