Maximizing cluster performance with Intel Cluster Studio XE

Getting the maximum performance in distributed clusters
Intel Cluster Studio XE
Werner Krotz-Vogel
Development Products Division
Software and Services Group
May 2014

Intel® Software Conference 2014
Agenda
Performance Tuning Methodology Overview
Quick overview of Intel® Trace Analyzer and Collector
What’s new in 2015 beta
Quick overview of Intel® VTune™ Amplifier XE
What’s new in 2015 beta
Performance Tuning Methodology using ITAC and VTune™ Amplifier XE
Demonstrated on Poisson Example
MPI 3.0 Support with Intel® MPI
Summary
2

Copyright© 2014, Intel Corporation. All rights reserved. *Other brands and names are the property of their respective owners.
Performance Tuning Methodology
using ITAC and VTune™ Amplifier XE
Step 1
•Cluster Level Analysis &
Algorithmic Tuning
Step 2
•Run-time Analysis & Tuning
Step 3
•Intra-Node and Single Node
Level Analysis

Intel® Trace Analyzer and Collector 8.1
What’s new?
4

Intel® Trace Analyzer and Collector 8.1 Update 3
What’s New
Fresh look-and-feel to the Intel® Trace
Analyzer Graphical Interface
 New toolbars, icons, and dialogs for more
streamlined analysis flow
 Addition of Welcome Page and easy access to
past projects
Support of Dynamic Profiling Tool
Command
 MPI_PControl supported
Support for MPI 2.x Standard
New GUI-based installer on Linux*
5
Compile
rs &
Libraries

Intel® MPI Library 5.0 Beta
Key Features
6

Initial MPI-3.0 Support
 Non-blocking and Sparse Collectives
 Fast Remote Memory Acess (RMA)
 Large buffer support (e.g. > 2GB) via mpi_count derived type
ABI compatibility with existing Intel® MPI Library and other MPICH*-based
applications
7
What’s New in Intel MPI Library 5.0 Beta

Support for the latest MPI-3.0 features
8
Use non-blocking collectives for a
complete comm/comp overlap
More efficient one-sided
communication via new Fast
Remote Memory Access
functionality
// Start synchronization
MPI_Ibarrier(comm, &req);
// Do extra computation
…
// Complete synchronization
MPI_Test(&req, …);
Example (C)

Intel® Trace Analyzer and Collector 9.0 Beta
Key Features
9

What’s New in Intel Trace Analyzer and Collector 9.0 Beta
10
Automatic Performance
Assistant
 Detect common MPI
performance issues
 Automated tips on potential
solutions

Intel® Trace Analyzer and Collector
Optimize MPI Communications (part of Intel® Cluster Studio XE)
 Visually understand parallel application
behavior
 Communications Patterns
 Hotspots
 Load Balance
 MPI Checking
 Detect Deadlocks
 Data Corruption
 Errors in Parameters, Data Types, etc
11
Intel®
ITAC
Processes
Year
0
1000
2000
3000
4000
5000
6000
7000
2010 2011 2012
Intel® Trace Analyzer and
Collector (processes)

ITAC 9.0: What’s New
12
• Collection
 Full MPI-3 support
 New mpirun options to customize collection
 Experimental TIME-WINDOWS support
 System calls profiling
• Analysis
 New Performance Assistant
 Visual appearance enhancement
 New Summary Page
• New tutorials

New mpirun data collection keys
13
Reduce a trace file size or a number of Message Checker reports (supported only at
runtime with Hydra process manager):
• -trace-collectives: collect info only about Collective operations
• -trace-pt2pt: collect info only about Point-to-Point operations
Example:
$ [mpirun|mpiexec] -trace-pt2pt –n 4 ./myApp

System calls profiling (1|2)
14
Linux* only. Capability to trace the following system calls:
access clearerr close creat
dup dup2 fclose fdopen
feof ferror fflush fgetc
fgetpos fgets fileno fopen
fprintf fputc fputs fread
freopen fseek fsetpos ftell
fwrite getc getchar gets
lseek lseek64 mkfifo perror
pipe poll printf putc
putchar puts read readv
remove rename rewind setbuf
setvbuf sync tmpfile tmpnam
umask ungetc vfprintf vprintf
write writev

System calls profiling (2|2)
15
To turn on system calls collection add any of the following lines into ITC configuration file:
• To collect all system calls:
ACTIVITY SYSTEM on
• To collect an exact function:
STATE SYSTEM:<func_name> ON
View system calls using ITA (new Group SYSTEM, can be expanded in an ordinary way):

New Summary Page
16
At-a-glance view on MPI activity and hints on how to start the analysis of the application:

New Performance Assistant
17
Automatic highlights of performance issues, both in GUI and CLI.
Currently 4 types of issues are supported, see screenshots:

Intel® VTune™ Amplifier XE 2013
Key Features
18

Intel® VTune™ Amplifier XE
Tune Applications for Scalable Multicore Performance
Fast, Accurate Performance Profiles
 Hotspot (Statistical call tree)
 Call counts (Statistical)
 Hardware-Event Sampling
Thread Profiling
 Visualize thread interactions on timeline
 Balance workloads
Easy set-up
 Pre-defined performance profiles
 Use a normal production build
Find Answers Fast
 Filter extraneous data
 View results on the source / assembly
Compatible
 Microsoft, GCC, Intel compilers
 C/C++, Fortran, Assembly, .NET, Java
 Latest Intel® processors
and compatible processors1
Windows or Linux
 Visual Studio Integration (Windows)
 Standalone user i/f and command line
 32 and 64-bit
19
1 IA32 and Intel® 64 architectures.
Many features work with compatible processors.
Event based sampling requires a genuine Intel® Processor.

What’s New in 2013 SP1?
Intel® VTune™ Amplifier XE
Intel Confidential
20
More Profiling Data
 Intel® Xeon Phi™ – memory and vectorization profiling
 Gen graphics tuning – GT event counting, offload, OpenCL*, …
Better Data Mining – Find Answers Faster
 Search added to all grids
 Timeline sorting, band height, time scale configuration
 Loop hierarchy, overhead and spin time metrics
 OpenMP* 4.0 – affinity controls, tasking and scalability analysis
Easier to Use
 Attach to a running Java process
 Contextual help for hardware events and performance metrics
 Easier generation of command line options from the user i/f
New OS & Processor Support
 Intel® Xeon Phi™, Haswell – Windows* & Linux*
 Windows 8 desktop and Visual Studio* 2012
 Collection on Windows UI and Windows Blue
 Latest Linux distributions
New since the first 2013 release. Some features released in earlier updates.

Intel® VTune™ Amplifier XE 2015 Beta
Key Features
21

 GPU analysis
 TSX analysis
 Remote collection in the GUI via ssh
 Mac OS* GUI data viewer (no collection)
 CSV import and custom collector support
 Timeline grouping
22
What’s New in Intel VTune™ Amplifier XE 2015 Beta

Step 1
Algorithmic Tuning
Step 2
Step 3
•Single Node Level Analysis
Global Analysis of the whole application gives first
indications of performance issues
• Run time and scaling analysis
• Message passing performance analysis on an inter/intra node level,
including finding of MPI hotspots
• Network Idealization that yields an imbalance diagram, providing
guidance on how to proceed
• Algorithmic/source code changes can be implemented for better
message passing practices or improving the load balance of the
application by:
• Fixing imbalances in communication patterns of MPI and non-
MPI routines.
• E.g: slow sequential I/O often causes imbalances.
• Removing unnecessary synchronization.
• E.g: message passing patterns using blocking send
and receive may cause a send/receive order that
increases wait times.
• This may be resolved by using non-blocking
MPI_Isend/MPI_Irecv pairs.

Step 1
Algorithmic Tuning
Step 2
Step 3
Intel MPI can be tuned without changing the
source code using:
• Environment variables for tuning of collective
operations, e.g., I_MPI_ADJUST_ALLREDUCE
• Environment variables for changing the message
passing characteristics, e.g.,
I_MPI_DAPL_DIRECT_COPY_THRESHOLD
• It is also possible to change the MPI
process/rank to node mapping for a better
inter/intra node communication balance

Step 1
Algorithmic Tuning
Step 2
Step 3
Single node tuning is necessary for serial and
parallel performance optimizations.
Single node tuning is important for improving overall
application scalability and reducing load imbalance.
Bandwidth analysis on the node is important for an
understanding of deficiencies in cluster level scaling.
Example: Conducting a hotspot analysis for each rank or
critical ranks identified in step 1 and 2.
The call stack information for a specific MPI routine may be
also helpful in refining of the analysis in Step 1.

Simple Scaling analysis
First step may be to just run the program for various number of processes [p] and
record timings: T[p]
Speedup S is defined as: S[p] = T[1]/T[p]
Efficiency E is defined as: E[p] = S[p]/p
An ideal parallel program will show:
S[p] = p and E[p] = 1

Example: Poisson solver on a square 3200x3200 computational
grid analysis
Poisson solver: simple implementation of Poisson solver: e.g. heat equation
• The 3200x3200 grid points can be distributed to MPI ranks using
• 2D process grid, e.g., in the case of 4 ranks, one can use 2 rows x 2 columns of processes
• 1D distribution with 4 rows x 1 column or 1 row x 4 columns
0 1 2 3
3200x800
local grid
points per
MPI rank
2 3
0 1
1600x1600
local grid
points per
MPI rank
Benchmark Environment
Intel® Xeon® E5 v2 processors (Ivy Town) with 12 cores. Frequency:
2 processors per node (24 cores per node)
Mellanox QDR Infiniband
Operating system: RedHat EL 6.1
Intel® MPI 4.1

Example: Poisson solver on a square 3200x3200 computational
grid analysis
Analysis of the application with different
numbers of processes (p)
Speed-up: S[p] = T[1]/T[p]
Parallel Efficiency: E[p] = S[p]/p
The speedup curves for the 2D quadratic
and 1D process (1D 1xN and 1D Nx1)
grids show some differences in scaling

Step 1- Cluster Level Analysis
using Intel Trace Analyzer and Collector
The MPI communication and
compute performance breakdown
of total run time
T[p] = T_comp[p] + T_mpi[p]
can be accessed through the trace
analyzer’s Function Profile (Intel®
Trace Analyzer displays the
Functions Profile Chart when
opening a trace file).
The trace file can be generated by
adding the flag “-trace” to the
mpirun or mpiexec.hydra command
The trace analyzer API
was used to time just 100
of 1653 iterations.
VT_API is paused time
Timing is accumulated
over Ranks. Application
time is T_comp
This column is the average
time per process. It can be
added by right click and
Function Profile Settings
Intel Trace Analyzer and Collector Flat Profile

Example - measuring MPI times with ITAC
Function Profile
MPI Breakdown of a
real Application
(VASP).
All MPI functions
are listed and may
be sorted by a click
on the top of each
column

Parallel efficiency can be calculated and
plotted for the compute time of the
application separately.
• Parallel Efficiency: E[p] = S[p]/p
• One can see that MPI Time is insignificant up
to 48 cores (the equivalent of two nodes).
• Above 96 ranks (4 nodes), pure computational
application performance also yields super
linear scaling.
• However, at the same data point, at around
96 ranks, MPI time becomes the main reason
for low efficiency

Message Passing Profile – 2D case (48x32)
• Message Passing Profile is a display of various
characteristics of message passing in a
sender/receiver matrix that can be obtained
through Charts-> Message Profile Chart.
• Dealing with 1536 ranks generates a huge matrix
we may fuse all ranks for each node: Advanced->
Process Aggregation-> All Nodes.
• The diagonal now shows the intra-node
performance characteristics while the off
diagonals show the inter-node statistics. Without
Process Aggregation the diagonal will be only
filled if we send messages from rank n to the
same rank n which is usually not a good idea

Message Passing Profile –1D case
(1535x1)
• Have much fewer massages in the 1D case
that are much larger, which leads to a much
higher average transfer rate.
• In the 1D case we also transfer a much
larger amount of data.

Algorithm and Network evaluation
ITAC shows timing of all MPI routines used by a program
The timing of these routines may be due to network transfer times caused by
bandwidth limitations
The other possibility are waiting times caused by the algorithm: load imbalance or
dependencies

A simple Network Model
The most simple network model defines latency
T_trans[V] = L + (1/BW)*V
• Latency L = transfer time for 0 byte message
• Bandwidth BW = transfer rate for (asymptotically) large messages.
• The transfer time is (V = Message Volume)

Ideal Network Simulator
It is extremely complicated to simulate a realistic network!
An extreme case – the ideal network – may be simulated by setting all transfer times
to 0. This would mean L = 0 and BW = ∞ for the simple model
ITAC offers an ideal network simulation with transfer times set to 0. Compute times
(non MPI) will stay the same

Idealization of network and the Load
Imbalance Diagram
• Employing the ideal network simulator (invoked
through the Advanced->Idealization menu)
allows us to separate network stack performance
impact on total MPI performance from algorithmic
inefficiencies like imbalance and dependencies.
• A simple network model for the transfer time as
a function of message volume V is
T_trans[V] = L + (1/BW)*V
• L is latency, defined as the time needed to transfer a 0 byte
message. BW is the transfer rate for asymptotically large
messages

Step 2- Runtime Analysis
It is possible to improve MPI performance without changing the source code
This can be done by using Intel MPI environment variables or by changing the
process mapping of ranks to compute nodes.
Process to node mapping can be altered by advanced methodologies like machine- or
configuration files or by reordering the ranks inside of a communicator
One option is start the tuning by concentrating on global operations
Set the environment variable I_MPI_DEBUG = 5.
 Prints valuable information about used variables, network fabrics and process placement.
 Setting I_MPI_DEBUG to 6 will further reveal the default algorithms for collective algorithms

Step 2- Runtime Analysis
Intel MPI reference guide reveals 8
different algorithms for MPI_Allreduce
• The algorithm can easily be changed
by setting the environment variable
I_MPI_ADJUST_ALLREDUCE to an
integer value in the range of 1-8

Step 3 - Analyzing Intra-Node performance
using Intel® VTune™ Amplifier XE
Intel Trace Analyzer and Collector is not
sufficient for hybrid applications due to its
primarily focus on MPI performance
• Hybrid codes that combine parallel MPI
processes with threading for a more
efficient exploitation of computing
resources
$> mpirun –n N amplxe-cl –result-dir
hotspots_N –collect hotspots -- poisson.x
ITAC analysis showed us which MPI functions are the hotspots.
But which MPI function?
MPI_Waitall function actually has the largest contribution to the application run time.

Hotspot
Functions

Hotspot MPI/System
functions

Callstack #3
Callstack #2
Callstack #1

using Intel® VTune™ Amplifier XE – Bandwidth Analysis
export VTUNE_COLLECT=snb-bandwidth
export VTUNE_COLLECT=hotspots
export VTUNE_FLAGS=-start-paused
mpiexec.hydra $MPI_FLAGS -n 1 amplxe-cl
$VTUNE_FLAGS --result-dir ${VTUNE_COLLECT}_$1 -
-collect $VTUNE_COLLECT

using Intel® VTune™ Amplifier XE - Bandwidth Analysis

8.908
61.836
83.548 86.569
80.717
30.392
15.839
0
0.2
0.4
0.6
0.8
1
1.2
1.4
0
10
20
30
40
50
60
70
80
90
100
1 6 12 24 48 72 96
ParallelEfficiency
Bandwidthononenode,GB/s
Number of ranks
Bandwidth vs. Parallel Efficiency on a first node
Bandwidth, GB/s Parallel Efficiency

MPI 3.0 Support with Intel® MPI
Intel® MPI Library 5.0 and Intel® Trace Analyzer and Collector 9.0

How do you spell MPI?
A de facto standard for communicating
between processes of a parallel program on
a distributed memory system
 Standardized
 Supported on almost all platforms
 Portable
 No need to modify your code when
porting
 Performance opportunities
 Vendor MPIs can exploit native
hardware features
 Functionality
 Over 125 routines defined by a
committee
49
#include "mpi.h“
int main(argc,argv){
MPI_Init(&argc,&argv);
MPI_Comm_size(MPI_COMM_WORLD,&numtasks);
MPI_Comm_rank(MPI_COMM_WORLD,&rank);
printf ("Number of tasks= %d
My rank= %dn",ntasks,rank);
/******* do some work *******/
MPI_Finalize();
}
Example (C)
MPI include file
Initialize MPI
environment
Terminate MPI
environment
Do work and
make MPI calls

What is in MPI-3?
Topic Motivation Main Result
Collective Operations Collective performance Non-Blocking & Sparse Collectives
Remote Memory Access Cache coherence, PGAS support Fast RMA
Backward Compatibility Buffers > 2 GB
Large buffer support, const
buffers
Fortran Bindings Fortran 2008
Fortran 2008 bindings
Removed C++ bindings
Tools Support PMPI Limitations MPIT Interface
Hybrid Programming Core count growth
MPI_Mprobe, shared memory
windows
Fault Tolerance Node count growth None. Next time?
50

I want a complete comm/comp overlap
Problem
 Computation/communication overlap is
not possible with the blocking collective
operations
Solution: Non-blocking Collectives
 Add non-blocking equivalents for
existing blocking collectives
 Do not mix non-blocking and blocking
collectives on different ranks in the
same operation
51
// Start synchronization
MPI_Ibarrier(comm, &req);
// Do extra computation
…
// Complete synchronization
MPI_Test(&req, …);
Example (C)

I have a sparse communication network
Problem
 Neighbor exchanges are poorly served
by the current collective operations
(memory and performance losses)
Solution: Sparse Collectives
 Add blocking and non-blocking
Allgather* and Alltoall*
collectives based on neighborhoods
52
call MPI_NEIGHBOR_ALLGATHER(&
& sendbuf, sendcount, sendtype,&
& recvbuf, recvcount, recvtype,&
& graph_comm, ierror)
Example (FORTRAN)

I want to use one-sided calls to reduce sync overhead
Problem
 MPI-2 one-sided operations are too
general to work efficiently on cache
coherent systems and compete with
PGAS languages
Solution: Fast Remote Memory Access
 Eliminate unnecessary overheads by
adding a ‘unified’ memory model
 Simplify usage model by supporting the
MPI_Request non-blocking call, extra
synchronization calls, relaxed
restrictions, shared memory, and much
more
53
call MPI_WIN_GET_ATTR(win, MPI_WIN_MODEL, &
memory_model, flag, ierror)
if (memory_model .eq. MPI_WIN_UNIFIED) then
! private and public copies coincide
Example (FORTRAN)

I’m sending *very* large messages
Problem
 Original MPI counts are limited to 2
Gigaunits, while applications want to
send much more
Solution: Large Buffer Support
 “Hide” the long counts inside the
derived MPI datatypes
 Add new datatype query calls to
manipulate long counts
54
// mpi_count may be, e.g.,
// 64-bit long
MPI_Get_elements_x(&status,
datatype, &mpi_count);
Example (FORTRAN)

None of these apply to me. What else you got?
I have a hybrid application
 Create a communicator inside a shared memory domain (intranode, via
MPI_Comm_split_type)
 Use the new MPI_Mprobe calls
I need to know what architecture I’m running on
 Predefined info object MPI_INFO_ENV allows for environment query
I’m using the C++ bindings
 Tough luck. C++ bindings have been removed from the standard.
55

Tell me more about this Intel® MPI Library
Optimized MPI application performance
 Application-specific tuning
 Automatic tuning
Lower Latency and Multi-vendor interoperability
 Optimized support for latest OFED* features
Faster MPI communication
 Optimized collectives
Sustainable scalability beyond 120K cores
 Native InfiniBand* interface allows for reduced
memory load and higher bandwidth
Simply and Accelerate Clusters
 Intel® Cluster Ready compliance
56
iWARPiWARP

Tell me more about this Intel® MPI Library
Optimized MPI application performance
 Application-specific tuning
 Automatic tuning
Lower Latency and Multi-vendor interoperability
 Optimized support for latest OFED* features
Faster MPI communication
 Optimized collectives
Sustainable scalability beyond 120K cores
 Native InfiniBand* interface allows for reduced
memory load and higher bandwidth
Simply and Accelerate Clusters
 Intel® Cluster Ready compliance
57
iWARPiWARP

Intel® MPI Library 5.0 & Intel® Trace Analyzer and Collector 9.0
Beta Nov 2013
 Non-blocking Collectives
 Fast RMA
 Large Counts
ABI compatibility with existing Intel® MPI
Library applications
58
Automatic Performance Assistant
 Detect common MPI performance issues
 Automated tips on potential solutions
Intel® MPI Library Intel® Trace Analyzer and Collector

What can I do?
Register for the Beta program (it’s free)
Start playing around with MPI-3.0
Come talk to us about it:
 Visit the Intel® Clusters and HPC Technology forums
 Check out the Intel® MPI Library product page (LEARN tab) for articles, examples, etc.
59
bit.ly/impi50-beta
software.intel.com/en-us/forums/intel-clusters-and-hpc-technologychnology
www.intel.com/go/mpi

Intel Technical Computing
Compute enables a New Scientific Method*
Technical computing and R&D workflow innovation
61
• Prediction
• Modeling & Simulation
• Experiment Refinement
• Physical
Prototyping
• Analysis
• Conclusion
• Refinement
• Physical
Prototyping
• Analysis
• Conclusion
• Refinement
• Hypothesis
• Hypothesis
1. Satava, Richard M. “The Scientific Method Is Dead-Long Live the (New) Scientific Method.” Journal of Surgical Innovation (June 2005).
• Prediction
Accelerates
the Method
Iterate

Intel® Technical Computing
Millions of Applications… Plus Yours
Delivering performance across generations and platforms
62
Today
Development
Tools
Performance/
Optimizations
Standards
Intel Architecture ecosystem: Increasing the return and longevity of your
application investment

Tablet
Desktop
Intel® Xeon ®
Workstation
Local Cluster
Computation Large Clusters
Common underlying architecture and software tools
scales investments across technical computing platforms
Intel® Technical Computing
The Right Tool for the Job: A Continuum of Computing
How do you get breakthroughs for your investment
63

Intel Confidential — Do Not Forward

Legal Disclaimer & Optimization Notice
INFORMATION IN THIS DOCUMENT IS PROVIDED “AS IS”. NO LICENSE, EXPRESS OR IMPLIED, BY ESTOPPEL OR OTHERWISE, TO
ANY INTELLECTUAL PROPERTY RIGHTS IS GRANTED BY THIS DOCUMENT. INTEL ASSUMES NO LIABILITY WHATSOEVER AND
INTEL DISCLAIMS ANY EXPRESS OR IMPLIED WARRANTY, RELATING TO THIS INFORMATION INCLUDING LIABILITY OR
WARRANTIES RELATING TO FITNESS FOR A PARTICULAR PURPOSE, MERCHANTABILITY, OR INFRINGEMENT OF ANY PATENT,
COPYRIGHT OR OTHER INTELLECTUAL PROPERTY RIGHT.
Software and workloads used in performance tests may have been optimized for performance only on Intel microprocessors.
Performance tests, such as SYSmark and MobileMark, are measured using specific computer systems, components, software,
operations and functions. Any change to any of those factors may cause the results to vary. You should consult other
information and performance tests to assist you in fully evaluating your contemplated purchases, including the performance of
that product when combined with other products.
Copyright © , Intel Corporation. All rights reserved. Intel, the Intel logo, Xeon, Xeon Phi, Core, VTune, and Cilk are trademarks of
Intel Corporation in the U.S. and other countries.
Optimization Notice
Intel’s compilers may or may not optimize to the same degree for non-Intel microprocessors for optimizations that are not
unique to Intel microprocessors. These optimizations include SSE2, SSE3, and SSSE3 instruction sets and other
optimizations. Intel does not guarantee the availability, functionality, or effectiveness of any optimization on
microprocessors not manufactured by Intel. Microprocessor-dependent optimizations in this product are intended for use
with Intel microprocessors. Certain optimizations not specific to Intel microarchitecture are reserved for Intel
microprocessors. Please refer to the applicable product User and Reference Guides for more information regarding the
specific instruction sets covered by this notice.
Notice revision #20110804

Maximizing cluster performance with Intel Cluster Studio XE

Recomendados

Recomendados

Más contenido relacionado

La actualidad más candente

La actualidad más candente (20)

Destacado

Destacado (18)

Similar a Maximizing cluster performance with Intel Cluster Studio XE

Similar a Maximizing cluster performance with Intel Cluster Studio XE (20)

Más de Intel Software Brasil

Más de Intel Software Brasil (16)

Último

Último (20)

Maximizing cluster performance with Intel Cluster Studio XE