Welcome to Hot Systems winter term 20002001

1
Welcome to Hot Systems winter term 2000/2001

• A course for absolute beginners and
• true experts in computer science and
• other more or less related sciences

2
Lecturers
Jochen Liedtke Uwe Dannowski Kevin
Elphinstone Espen Skoglund Gerd Liefländer
3
Overview

• General objectives of this course
• Introduction into topic of today
• Literature on the NANOS project
• People of the NANOS project
• Related research community
• Performance-Driven Processor Allocation

4
Objectives of the Course

• Hot Systems brand new, interesting work
• Insight into current research of systems
• Conference and technical papers
• Introduction to each specific problem
• Focus on the basic ideas and objectives
• No implementation details

5
Topic of Today

• Performance-Driven Processor Allocation
• by
• Julita Corbalan, Xavier Martorell, Jesús Labarta
• Departement dArquitectura de Computadores (DAC)
• Universtitat Politècnica de Catalunya (UPC)
• given at the OSDI 2000 in San Diègo, USA

6
Topic of Today

• Problem
• How to schedule parallel applications
• on a SMP (NUMA) efficiently?

7
Introduction to Topic of Today
Schedule?

• Problem
• How to schedule parallel applications
• on a SMP (NUMA) efficiently?

8
Introduction to Topic of Today
Schedule arrange something, i.e. determine
sequence of execution order

• Problem
• How to schedule parallel applications
• on a SMP (NUMA) efficiently?

9
Introduction to Topic of Today
Parallel Applications?
Schedule !

• Problem
• How to schedule parallel applications
• on a SMP (NUMA) efficiently?

10
Introduction to Topic of Today
Parallel Applications?
Schedule !

• Problem
• How to schedule parallel applications
• on a SMP (NUMA) efficiently?

SMP?
11
Introduction to Topic of Today
Parallel Applications?
Schedule !

• Problem
• How to schedule parallel applications
• on a SMP (NUMA) efficiently?

SMP shared memory multi-processor or symmetric
multi-processor
12
Introduction to Topic of Today
Parallel Applications?
Schedule !

• Problem
• How to schedule parallel applications
• on a SMP (NUMA) efficiently?

SMP !
Efficiently?
13
Introduction to Topic of Today
Parallel Applications?
Schedule !

• Problem
• How to schedule parallel applications
• on a SMP (NUMA) efficiently?

Efficiently (sub)optimal according to a
performance measure with low(est) overhead
SMP !!
14
Scheduling Problem
Multi-threaded application

Resources
How to schedule?
T0 T1
CPU 0 CPU 1 CPU 2 CPU 3
T2 T3
T4 T5
Scheduling parameters 1. Number of processors
being involved 2. Precedence relation 3.
Communication costs 4. Memory access time
because of (NUMA)
15
Scheduling Problem
Multi-threaded application

Resources
How to schedule?
T0 T1
CPU 0 CPU 1 CPU 2 CPU 3
T2 T3
T4 T5
x ? 1, p static/dynamic assignment Pros and
Cons? If x too large then danger of idle CPUs if
x too small then long execution times
Scheduling parameters 1. Number of processors
being involved 2. Precedence relation 3.
Communication costs 4. Memory access time
because of (NUMA)
16
Scheduling Problem
Multi-threaded application

Resources
How to schedule?
T0 T1
CPU 0 CPU 1 CPU 2 CPU 3
T2 T3
T4 T5
x ? 1, p static/dynamic assignment in this
talk Pros and Cons? If x too large then danger
of idle CPUs if x too small then long execution
times
Scheduling parameters 1. Number of processors
being involved 2. Precedence relation 3.
Communication costs 4. Memory access time
because of (NUMA)
17
Scheduling Problem
Multi-threaded application

Resources
How to schedule?
T0 T1
CPU 0 CPU 1 CPU 2 CPU 3
T2 T3
T4 T5
Scheduling parameters 1. Number of processors
being involved 2. Precedence relation 3.
Communication costs 4. Memory access time
because of (NUMA)
Whether Tj must finish before Ti can execute
18
Scheduling Problem
Multi-threaded application

Resources
How to schedule?
T0 T1
CPU 0 CPU 1 CPU 2 CPU 3
T2 T3
T4 T5
Scheduling parameters 1. Number of processors
being involved 2. Precedence relation 3.
Communication costs 4. Memory access time
because of (NUMA)
Threads interacting heavily should run on
neighbored processors !!
19
Scheduling Problem
Multi-threaded application

Resources
How to schedule?
T0 T1
CPU 0 CPU 1 CPU 2 CPU 3
T2 T3
T4 T5
Scheduling parameters 1. Number of processors
being involved 2. Precedence relation 3.
Communication costs 4. Memory access time
because of (NUMA)
Place the shared memory of cooperating threads
close together !
20
Scheduling Problem
Multi-threaded application

Resources
How to schedule?
T0 T1
CPU 0 CPU 1 CPU 2 CPU 3
T2 T3
T4 T5
Scheduling parameters 1. Number of processors
to be involved 2. Precedence relation 3.
Communication costs 4. Memory access time
because of (NUMA)
Non uniform memory access
21
Scheduling Problem

• Some still,open questions
• What kind of applications do we bear in mind?
• Who tells us how to decompose an
• application system into ngt1 threads?
• How can they work together?
• Which hard-/software constraints
• do we have to meet and when?

Multi-threaded application

Resources
How to schedule?
T0 T1
CPU 0 CPU 1 CPU 2 CPU 3
T2 T3
T4 T5
Scheduling parameters 1. Number of processors
to be involved 2. Precedence relation 3.
Communication costs 4. Memory access time
because of (NUMA)
22
Parallel Application

• Numerical problems
• Iterative algorithms
• Simulations

This playground is very restricted! More or less
scientific computing, e.g. weather-forecasting
etc. Why do we deal with anyhow?
23
Parallel Application

• Numerical problems
• Iterative algorithms
• Simulations

1. Better start with simple, regular patterns to
obtain understandable results 2. These
applications have long execution times, so ...
3. These applications run quite often, though
...
24
Parallel Application

• Numerical problems
• Iterative algorithms
• Simulations
• Summary
• All with a more or less regular execution
  pattern

25
Structure of Parallel Application

• initialize some data, i.e. info(1)
• for k1(1)X do
• prepare something sequential
• for i1(1)n do
• for j1(1)n do
• calculate info(k1) with info(k)
• end inner loop
• end outer loop
• synchronize all outer loops
• end algorithm loop
• prepare info(n) to be printed

26
Structure of Parallel Application
?
27
Decomposing Impetus

• Application programmer
• Compiler
• Runtime system

28
Application Characteristics

• Maximum speedup
• Average parallelism
• Size of working set
• Communication costs
• How to achieve those characteristics?

29
Application Characteristics
Both methods are either time consuming and/or
somehow inaccurate.

• Maximum speedup
• Average parallelism
• Size of working set
• Communication costs
• 1. User performs several executions with
  different scenarios thus
• collecting several measurements.
• 2. Define an analytic job model characterizing
  the application by a
• set of parameters. This information is an a
  priori input to each
• execution of the application.

30
Application Characteristics

• Maximum speedup
• Average parallelism
• Size of working set
• Communication costs
• Solution by Self-Analyzing
• Calculate job characteristics at run time
• and use them for processor scheduling

31
Conference Papers(1) "A Library Implementation
of the Nano-Threads Programming Model" Xavier
Martorell, Jesus Labarta, Nacho Navarro and
Eduard Ayguade Proc. of 2th International
Euro-Par Conference, August 1996. "Analysis of
Several Scheduling Algorithms under the
Nano-Threads Programming Model", Xavier
Martorell, Jesus Labarta, Nacho Navarro and
Eduard Ayguade Proc. of 11th International
Parallel Processing Symposium, April 1997.
"Exploiting Parallelism Through Directives on
the Nano-Threads Programming Model" Eduard
Ayguade, Xavier Martorell, Jesus Labarta, Marc
Gonzalez and Nacho Navarro Proc. of 10th Intern.
Workshop on Languages a. Compilers for Parallel
Computing, August 1997. "Efficient Runtime
Thread Management for the Nano-Threads
Programming Model" D.S. Nikolopoulos, E.D.
Polychronopoulos and T.S. Papatheodorou Proc. of
Workshop on Runtime Systems for Parallel
Programming, April 1998. "Kernel-level
Scheduling for the Nano-Threads Programming
Model" E. D. Polychronopoulos, X. Martorell, D.
S. Nikolopoulos, J. Labarta, T. S. Papatheodorou
and N. Navarro, Proc. of 12th ACM Intern.
Conference on Supercomputing, July 1998.
32
Conference Papers(2) "Enhancing the Performance
of Autoscheduling in Distributed Shared Memory
Multiprocessors" D.S. Nikolopoulos, E. D.
Polychronopoulos and T. S. Papatheodorou Proc.
of 4th International Euro-Par Conference,
September 1998. "HTGviz A Graphics Tool for
the Synthesis of Automatic and User-driven
Program Parallelization in Compilation Process,
M. Giordano and M. Mango Furnari Proc. of 1st
International Symposium on High-Performance
Computing, May 1999. "A Quantitative Evaluation
of Synchronization Algorithms and Disciplines on
ccNUMA Systems The Case of the SGI Origin2000,
D.S. Nikolopoulos and T.S. Papatheodorou Proc.
of 1999 International Conference on
Supercomputing, June 1999. "Thread Fork/Join
Techniques for Multi-level Parallelism
Exploitation in NUMA Multiprocessors" Xavier
Martorell, Eduard Ayguade, Nacho Navarro, Julita
Corbalan, Marc Gonzalez and Jesus Labarta, Proc.
of 1999 International Conference on
Supercomputing, June 1999. "A Graphic
Parallelizing Environment for User-compiler
Interaction" M. Giordano and M. Mango Furnari
Proc. of 1999 International Conference on
Supercomputing, June 1999.
33
Conference Paper(3) "Fine-Grain and
Multiprogramming-Conscious Nanothreading with the
Solaris Operating System" D.S. Nikolopoulos,
E.D. Polychronopoulos and T.S. PapatheodorouProc.
of 1999 International Conference on Parallel and
Distributed Processing techniques and
Applications, June 28 - July 1 1999. "Achieving
Multiprogramming Scalability of Parallel Programs
on Intel SMP Platforms Nanothreading in the
Linux Kernel, D.S. Nikolopoulos, E.D.
Polychronopoulos, T.S. Papatheodorou, C.D.
Antonopoulos, I.E. Venetis and P.E. Hadjidoukas
Proc. of Parallel Computing'99, August 1999.
"An Efficient Kernel-Level Scheduling
Methodology for Multiprogrammed Shared Memory
Multiprocessors, E.D. Polychronopoulos, D.S.
Nikolopoulos, T.S. Papatheodorou, X. Martorell,
J. Labarta and N. Navarro, Proc. of 12th
International Conference on Parallel and
Distributed Computing Systems, August 1999.
"Scheduling User-Level Threads on Distributed
Shared Memory Multiprocessors" E.D.
Polychronopoulos, and T.S. Papatheodorou5th
EuroPar Conference, September 1999. "Improving
the Performance of Multiprogrammed Parallel
Workloads in Origin2000 Systems" Xavier
Martorell, Eduard Ayguade, Jesus Labarta and
Nacho Navarro 5th European Cray/SGI MPP
Workshop, September 1999.
34
Conference Paper(4) "Exploiting Multiple Levels
of Parallelism in OpenMP A Case Study" Eduard
Ayguade, Xavier Martorell, Jesus Labarta, Marc
Gonzalez and Nacho Navarro Proc. of the 1999
International Conference on Parallel Processing,
September 1999. "NanosCompiler A Research
Platform for OpenMP Extensions" Eduard Ayguade,
Marc Gonzalez, Jesus Labarta, Xavier Martorell,
Nacho Navarro and Jose Oliver, Proc. of the 1st
European Workshop on OpenMP, October 1999.
"Code Generation and Run-time Support for
Multi-level Parallelism Exploitation" Marc
Gonzalez, Xavier Martorell, Jose Oliver, Eduard
Ayguade and Jesus LabartaProc. of 8th
International Workshop on Compilers for Parallel
Computing, January 2000. "Fast Synchronization
on Scalable Cache-Coherent Multiprocessors Using
Hybrid Primitives" D. S. Nikolopoulos and T. S.
PapatheodorouIEEE 2000 International Parallel
and Distributed Processing Symposium, May 2000.
"Applying Interposition Techniques for
Performance Analysis of OpenMP Applications" M.
Gonzalez, X. Martorell, J. Oliver, A. Serra, E.
Ayguade, J. Labarta and N. NavarroIEEE 2000
International Parallel and Distributed Processing
Symposium, May 2000.
35
Conference Paper(5) A Tool to Schedule Parallel
Applications on Multiprocessors The NANOS
CPUManagerXavier Martorell, Julita Corbalán,
Dimitrios Nikolopoulos, José I. Navarro,
Eleftherios Polychronopoulos, Theodore
Papatheodorou and Jesús Labarta In 6th Annual
Workshop on Job Scheduling Strategies for
Parallel Processing, May 2000. Performance-Drive
n Processor Allocation" Julita Corbalan, Xavier
Martorell, Jesus Labarta, Proc. of OSDI 2000,
October 2000. The NANOS Resource Management
System Xavier Martorell, Julita Corbalan, Nachos
Navarro, Jesus Labarta, Poster Session on OSDI
2000, October 2000. Summary 23 Conference
Papers by 16 authors within 5 years! Thats
well done!
36
Technical Papers of Julita Corbalan
(1) Improving Processor Allocation through
Run-Time Measured EfficiencyJulita Corbalán,
Jesus Labarta Submitted to International Parallel
and Distributed Processing Symposium, April
2001 A Dynamic Periodicity Detector
Application to Speedup Computation" Felix
Freitag, Julita Corbalan, Jesus Labarta,
Submitted to International Parallel and
Distributed Processing Symposium, April 2001. A
Processor Scheduler The CPUManager Julita
Corbalán, Jesus Labarta, Xavier Martorell,
UPC-DAC-1999-69 Dynamic Speeup Calculation
through Self-Analysis Julita Corbalán, Jesus
Labarta, UPC-DAC-1999-43 The NANOS Environment
User Guide E. Ayguade, C.R. Calidonna, Julita
Corbalan, M. Giordano, M. Gonzales, H.C. Hoppe,
Jesus Labarta, M. Furnari, Xavier Martorell, N.
Navarro, D.S. Nikolopolous, J. Oliver, T.S.
Papatheodorou, E.D. Polychronopolous,
UPC-DAC-1999-42
37
Technical Papers of Julita Corbalan
(1) Improving Processor Allocation through
Run-Time Measured EfficiencyJulita Corbalán,
Jesus Labarta Submitted to International Parallel
and Distributed Processing Symposium, April
2001 A Dynamic Periodicity Detector
Application to Speedup Computation" Felix
Freitag, Julita Corbalan, Jesus Labarta,
Submitted to International Parallel and
Distributed Processing Symposium, April 2001. A
Processor Scheduler The CPUManager Julita
Corbalán, Jesus Labarta, Xavier Martorell,
UPC-DAC-1999-69 Dynamic Speeup Calculation
through Self-Analysis Julita Corbalán, Jesus
Labarta, UPC-DAC-1999-43 The NANOS Environment
User Guide E. Ayguade, C.R. Calidonna, Julita
Corbalan, M. Giordano, M. Gonzales, H.C. Hoppe,
Jesus Labarta, M. Furnari, Xavier Martorell, N.
Navarro, D.S. Nikolopolous, J. Oliver, T.S.
Papatheodorou, E.D. Polychronopolous,
UPC-DAC-1999-42
38
Technical Papers of Julita Corbalan (2) NANOS
Effective Integration of Fine-grain Parallelism
Exploitation and Multiprogramming E. Ayguade,
C.R. Calidonna, Julita Corbalan, M. Giordano, M.
Gonzales, H.C. Hoppe, Jesus Labarta, M. Furnari,
Xavier Martorell, N. Navarro, D.S. Nikolopolous,
J. Oliver, T.S. Papatheodorou, E.D.
Polychronopolous, UPC-DAC-1999-42 Environment
and Benchmark Selection for the NANOS Project E.
Ayguade, Julita Corbalan, T. Cortes, M. Furnari,
M. Giordano, M. Gonzales, H.C. Hoppe, Jesus
Labarta, Xavier Martorell, R. Napolitano, N.
Navarro, D.S. Nikolopolous, T.S. Papatheodorou,
E.D. Polychronopolous, UPC-DAC-1998-14 Manual
Parallelization of Applications using the NANOS
Programming Model E. Ayguade, Julita Corbalan,
M. Furnari, M. Giordano, M. Gonzales, H.C. Hoppe,
Jesus Labarta, Xavier Martorell, N. Navarro,
D.S. Nikolopolous, T.S. Papatheodorou, E.D.
Polychronopolous, UPC-DAC-1998-14 and 2
additional Spanish papers.
39
NANOS Project Members
CEPBA-UPC (Project Coordinator) Universitat
Politècnica de Catalunya European Center of
Parallelism Barcelona, Spain Prof. Eduard Ayguadé
LHPCA-UP University of Patras Greece
CNR Consiglio Nationalle delle Richerce Istituto
di Cibernetica, Napoli, Italy
Pallas GmbHHermülheimer Str.10D-50321
BrühlGermany
40
Some NANOS Project People
Eduard Ayguadé
Julita Corbalan