|
|
 |
|
Overview |
|
|
|
|
Intel® Cluster Tools assist developers and managers of distributed
systems to obtain the best performance.
Intel® Trace Collector1 brings the advantages of
event-based tracing to applications in a low-overhead tracing library.
It supports MPI, Java*, and multi-threaded processes with and without
MPI. The tool is completely thread safe, allowing tracing of multi-threaded
MPI applications. Automatic function profiling is also now supported
on all platforms when the GNU Compiler Collection is used to compile
C or Fortran source code. Intel® Trace Analyzer1
and Intel Trace Collector support IA-32 and Intel® Itanium®
architectures with Linux* (MPICH). Intel Trace Collector supports
the LAM MPI implementation on IA-32 architectures and HPUX* (HP-MPI)
and SGI Altix* (SGI-MPI) systems on Intel Itanium 2 microarchitectures.
Intel Trace Analyzer displays the data produced by Intel Trace Collector
|
|
|
|
|
Features and Benefits
 |
Profiling library records distributed, event-based
trace data |
|
|
Low-overhead and compact data representation
provides structured trace file format (STF) |
|
|
Completely thread safe to allow tracing of
multi-threaded message passing interface (MPI) applications |
|
|
Easy-to-use application program interface
(API) |
|
|
Low intrusion instrumentation for MPI, Java*,
or multi-threaded applications |
|
|
Complements Intel’s offerings designed
for developing high-performance software including compilers,
libraries, and Intel® VTune™ Performance Analyzers |
|
|
Automatic function profiling supported on
all platforms when the GNU Compiler Collection is used to
compile C or Fortran source code |
|
|
Global Array programming model allows tracing and performance
analysis of applications |
|
|
|
Graphically Analyze Runtime Event Traces
|
|
Event-based tracing tool — Successful use of Intel Trace
Collector for MPI performance analysis demonstrates the advantages
of event-based tracing compared to exclusively statistical
approaches. The timeline display helps developers visualize
the concurrent behavior of parallel applications. Statistics
can be calculated on demand for given time intervals and specific
processes. The tool also brings the advantages of event-based
tracing to non-MPI applications such as Java processes and
multi-threaded processes without MPI. Java function profiling
is possible without recompiling the Java classes. In addition,
the traditional Virtual Terminal application programming interface
(VT API) is also available to instrument Java source code.
|
Ease of use — Intel Trace Collector works as an
add-on to existing MPI implementations. In most cases,
using the tool requires relinking with Intel Trace
Collector library. If application-defined events must
be recorded, a recompile may be necessary. The low-overhead
instrumentation limits the perturbation of the application
execution, assuring that the runtime behavior remains
fundamentally unaltered. Experience shows that lessons
learned from studying the traces of a run with Intel
Trace Collector invariably lead to improved performance
of non-instrumented runs. |
|
|
Structured trace file format — Intel Trace Collector
introduces a new file format for trace data: the Structured
Trace Format (STF). This file format was designed from
the ground up for scalability and compact data representation.
Files can be written in parallel, thus generating trace
files faster, and allows random access to portions of
a trace, making it suitable for analysis of traces too
large for Intel Trace Analyzer. |
|
|
Scalability and memory handling — Reduction of
large memory consumption, often associated with event-based
tracing, is addressed by the filtering and memory-handling
capabilities of Intel Trace Collector. For deep function
call stacks, folding avoids superfluous details by logging
the first call to a system function and then hiding
all internal functions below this call. Trace data is
cached in memory to reduce runtime overhead, but can
be written to a cache file in the background without
blocking the application. Counters can be used to monitor
Intel Trace Collector's memory handling. If that still
interferes with the application or simply produces unneeded
data, Intel Trace Collector can be used in a purely
statistical mode. In this mode, statistics regarding
function calls, messages, and collective operations
are calculated by Intel Trace Collector at runtime and
stored without event data. Intel Trace Analyzer can
display these statistics instantaneously without having
to load any trace data. |
|
|
Completely thread safe — Intel Trace Collector
allows tracing of multi-threaded MPI applications. The
tool brings the advantages of event-based tracing to
non-MPI applications by supporting Java processes and
multi-threaded processes without MPI. Java function
profiling is possible without recompiling the Java classes.
In addition, the traditional VT API is available to
instrument Java source code. Automatic function profiling
is supported on all platforms when the GNU Compiler
Collection is used to compile C or Fortran source code.
|
|
|
|
|
PERFORMANCE |
|
|
Processor and Data Scalability |
|
|
Intel Trace Collector easily scales up to several hundred processors
due to a hierarchical approach to address data scalability.
|
|
|
COMPATIBILITY |
|
|
Diverse Hardware and Software Support |
|
|
Intel Trace Collector supports IA-32 and Itanium architectures with
Linux. MPI implementations such as MPICH, LAM MPI, ScaMPI, and Myrinet*
MPI are supported on IA-32 architectures. On Itanium architectures,
MPICH, HP-UX (HP-MPI), and SGI Altix (SGI-MPI) systems are supported.
Automatic function profiling is supported on all platforms when
the GNU Compiler Collection is used to compile C or Fortran source
code.
|
|
|
SUPPORT |
|
|
Intel® Premier Support |
|
|
Every purchase of an Intel® Software Development Product includes
a year of support services, which provides access to Intel®
Premier Support and all product updates during that time. Intel
Premier Support gives you online access to technical notes, application
notes, and documentation. Install the product, and then register
to get support and product update information.
|
|
|
REQUIREMENTS
Hardware
|
|
|
|
|
Intel® Pentium® 4, Intel® Xeon™, or newer
processor |
|
|
512 MB RAM |
|
|
1 GB disk space |
or
|
|
Intel® Itanium® 2 processor |
|
|
1 GB RAM |
|
|
1 GB disk space |
|
|
|
|
Software
Supports IA-32 and Itanium architecture-based systems with Linux*.
MPI implementations such as MPICH, LAM MPI, ScaMPI, and Myrinet*
MPI are supported on IA-32 architectures. On Itanium architecture-based
systems, MPICH, HP-UX*, and SGI Altix* systems are supported. |
|
|
|
Note: Intel Trace Collector Parallel
Performance Analysis Library is a machine (cluster) bound license.
This means for every machine (cluster), you need a separate license. Licensing
of Intel Trace Collector1 Parallel Performance Analysis Libraries
restricts the maximum number of MPI processes used in a parallel
program to no more than the total number of processors used, regardless
of the number of MPI processes executed per processor. |
|
|
|
|
|
|
|
| |
|
|
|
|
|
