ompi/README at master · brminich/ompi

Repository navigation

History

100

101

102

103

104

105

106

107

108

109

110

111

112

113

114

115

116

117

118

119

120

121

122

123

124

125

126

127

128

129

130

131

132

133

134

135

136

137

138

139

140

141

142

143

144

145

146

147

148

149

150

151

152

153

154

155

156

157

158

159

160

161

162

163

164

165

166

167

168

169

170

171

172

173

174

175

176

177

178

179

180

181

182

183

184

185

186

187

188

189

190

191

192

193

194

195

196

197

198

199

200

201

202

203

204

205

206

207

208

209

210

211

212

213

214

215

216

217

218

219

220

221

222

223

224

225

226

227

228

229

230

231

232

233

234

235

236

237

238

239

240

241

242

243

244

245

246

247

248

249

250

251

252

253

254

255

256

257

258

259

260

261

262

263

264

265

266

267

268

269

270

271

272

273

274

275

276

277

278

279

280

281

282

283

284

285

286

287

288

289

290

291

292

293

294

295

296

297

298

299

300

301

302

303

304

305

306

307

308

309

310

311

312

313

314

315

316

317

318

319

320

321

322

323

324

325

326

327

328

329

330

331

332

333

334

335

336

337

338

339

340

341

342

343

344

345

346

347

348

349

350

351

352

353

354

355

356

357

358

359

360

361

362

363

364

365

366

367

368

369

370

371

372

373

374

375

376

377

378

379

380

381

382

383

384

385

386

387

388

389

390

391

392

393

394

395

396

397

398

399

400

401

402

403

404

405

406

407

408

409

410

411

412

413

414

415

416

417

418

419

420

421

422

423

424

425

426

427

428

429

430

431

432

433

434

435

436

437

438

439

440

441

442

443

444

445

446

447

448

449

450

451

452

453

454

455

456

457

458

459

460

461

462

463

464

465

466

467

468

469

470

471

472

473

474

475

476

477

478

479

480

481

482

483

484

485

486

487

488

489

490

491

492

493

494

495

496

497

498

499

500

501

502

503

504

505

506

507

508

509

510

511

512

513

514

515

516

517

518

519

520

521

522

523

524

525

526

527

528

529

530

531

532

533

534

535

536

537

538

539

540

541

542

543

544

545

546

547

548

549

550

551

552

553

554

555

556

557

558

559

560

561

562

563

564

565

566

567

568

569

570

571

572

573

574

575

576

577

578

579

580

581

582

583

584

585

586

587

588

589

590

591

592

593

594

595

596

597

598

599

600

601

602

603

604

605

606

607

608

609

610

611

612

613

614

615

616

617

618

619

620

621

622

623

624

625

626

627

628

629

630

631

632

633

634

635

636

637

638

639

640

641

642

643

644

645

646

647

648

649

650

651

652

653

654

655

656

657

658

659

660

661

662

663

664

665

666

667

668

669

670

671

672

673

674

675

676

677

678

679

680

681

682

683

684

685

686

687

688

689

690

691

692

693

694

695

696

697

698

699

700

701

702

703

704

705

706

707

708

709

710

711

712

713

714

715

716

717

718

719

720

721

722

723

724

725

726

727

728

729

730

731

732

733

734

735

736

737

738

739

740

741

742

743

744

745

746

747

748

749

750

751

752

753

754

755

756

757

758

759

760

761

762

763

764

765

766

767

768

769

770

771

772

773

774

775

776

777

778

779

780

781

782

783

784

785

786

787

788

789

790

791

792

793

794

795

796

797

798

799

800

801

802

803

804

805

806

807

808

809

810

811

812

813

814

815

816

817

818

819

820

821

822

823

824

825

826

827

828

829

830

831

832

833

834

835

836

837

838

839

840

841

842

843

844

845

846

847

848

849

850

851

852

853

854

855

856

857

858

859

860

861

862

863

864

865

866

867

868

869

870

871

872

873

874

875

876

877

878

879

880

881

882

883

884

885

886

887

888

889

890

891

892

893

894

895

896

897

898

899

900

901

902

903

904

905

906

907

908

909

910

911

912

913

914

915

916

917

918

919

920

921

922

923

924

925

926

927

928

929

930

931

932

933

934

935

936

937

938

939

940

941

942

943

944

945

946

947

948

949

950

951

952

953

954

955

956

957

958

959

960

961

962

963

964

965

966

967

968

969

970

971

972

973

974

975

976

977

978

979

980

981

982

983

984

985

986

987

988

989

990

991

992

993

994

995

996

997

998

999

1000

University Research and Technology

of Tennessee Research Foundation. All rights

reserved.

$COPYRIGHT$

Additional copyrights may follow

$HEADER$

===========================================================================

When submitting questions and problems, be sure to include as much

extra information as possible. This web page details all the

information that we request in order to provide assistance:

http://www.open-mpi.org/community/help/

The best way to report bugs, send comments, or ask questions is to

and developer-level questions; when in doubt, send to the user's

list):

users@open-mpi.org

devel@open-mpi.org

Because of spam, only subscribers are allowed to post to these lists

(ensure that you subscribe with and post from exactly the same e-mail

address -- joe@example.com is considered different than

joe@mycomputer.example.com!). Visit these pages to subscribe to the

lists:

http://www.open-mpi.org/mailman/listinfo.cgi/users

http://www.open-mpi.org/mailman/listinfo.cgi/devel

Thanks for your time.

===========================================================================

Much, much more information is also available in the Open MPI FAQ:

http://www.open-mpi.org/faq/

===========================================================================

The following abbreviated list of release notes applies to this code

base as of this writing (April 2015):

General notes

-------------

- Open MPI now includes two public software layers: MPI and OpenSHMEM.

Throughout this document, references to Open MPI implicitly include

both of these layers. When distinction between these two layers is

necessary, we will reference them as the "MPI" and "OSHMEM" layers

respectively.

- OpenSHMEM is a collaborative effort between academia, industry, and

the U.S. Government to create a specification for a standardized API

for parallel programming in the Partitioned Global Address Space

(PGAS). For more information about the OpenSHMEM project, including

access to the current OpenSHMEM specification, please visit:

http://openshmem.org/

This OpenSHMEM implementation is provided on an experimental basis;

it has been lightly tested and will only work in Linux environments.

Although this implementation attempts to be portable to multiple

different environments and networks, it is still new and will likely

experience growing pains typical of any new software package.

End-user feedback is greatly appreciated.

This implementation will currently most likely provide optimal

performance on Mellanox hardware and software stacks. Overall

performance is expected to improve as other network vendors and/or

institutions contribute platform specific optimizations.

See below for details on how to enable the OpenSHMEM implementation.

- Open MPI includes support for a wide variety of supplemental

hardware and software package. When configuring Open MPI, you may

need to supply additional flags to the "configure" script in order

to tell Open MPI where the header files, libraries, and any other

required files are located. As such, running "configure" by itself

may not include support for all the devices (etc.) that you expect,

especially if their support headers / libraries are installed in

non-standard locations. Network interconnects are an easy example

to discuss -- Libfabric and OpenFabrics networks, for example, both

have supplemental headers and libraries that must be found before

Open MPI can build support for them. You must specify where these

files are with the appropriate options to configure. See the

listing of configure command-line switches, below, for more details.

- The majority of Open MPI's documentation is here in this file, the

included man pages, and on the web site FAQ

(http://www.open-mpi.org/).

- Note that Open MPI documentation uses the word "component"

frequently; the word "plugin" is probably more familiar to most

users. As such, end users can probably completely substitute the

word "plugin" wherever you see "component" in our documentation.

For what it's worth, we use the word "component" for historical

reasons, mainly because it is part of our acronyms and internal API

function calls.

- The run-time systems that are currently supported are:

- rsh / ssh

- LoadLeveler

- PBS Pro, Torque

- Platform LSF (v7.0.2 and later)

- SLURM

- Cray XE, XC, and XK

- Oracle Grid Engine (OGE) 6.1, 6.2 and open source Grid Engine

- Systems that have been tested are:

- Linux (various flavors/distros), 32 bit, with gcc

- Linux (various flavors/distros), 64 bit (x86), with gcc, Absoft,

Intel, and Portland (*)

- OS X (10.6, 10.7, 10.8, 10.9, 10.10), 32 and 64 bit (x86_64), with

XCode and Absoft compilers (*)

(*) Be sure to read the Compiler Notes, below.

- Other systems have been lightly (but not fully tested):

- Cygwin 32 & 64 bit with gcc

- ARMv4, ARMv5, ARMv6, ARMv7, ARMv8

- Other 64 bit platforms (e.g., Linux on PPC64)

- Oracle Solaris 10 and 11, 32 and 64 bit (SPARC, i386, x86_64),

with Oracle Solaris Studio 12.2, 12.3, and 12.4

Compiler Notes

--------------

- Open MPI requires a C99-capable compiler to build.

- Mixing compilers from different vendors when building Open MPI

(e.g., using the C/C++ compiler from one vendor and the Fortran

compiler from a different vendor) has been successfully employed by

some Open MPI users (discussed on the Open MPI user's mailing list),

but such configurations are not tested and not documented. For

example, such configurations may require additional compiler /

linker flags to make Open MPI build properly.

- In general, the latest versions of compilers of a given vendor's

series have the least bugs. We have seen cases where Vendor XYZ's

compiler version A.B fails to compile Open MPI, but version A.C

(where C>B) works just fine. If you run into a compile failure, you

might want to double check that you have the latest bug fixes and

patches for your compiler.

- Users have reported issues with older versions of the Fortran PGI

compiler suite when using Open MPI's (non-default) --enable-debug

configure option. Per the above advice of using the most recent

version of a compiler series, the Open MPI team recommends using the

latest version of the PGI suite, and/or not using the --enable-debug

configure option. If it helps, here's what we have found with some

(not comprehensive) testing of various versions of the PGI compiler

suite:

pgi-8 : NO known good version with --enable-debug

pgi-9 : 9.0-4 known GOOD

pgi-10: 10.0-0 known GOOD

pgi-11: NO known good version with --enable-debug

pgi-12: 12.10 known BAD due to C99 compliance issues, and 12.8

and 12.9 both known BAD with --enable-debug

pgi-13: 13.10 known GOOD

- Similarly, there is a known Fortran PGI compiler issue with long

source directory path names that was resolved in 9.0-4 (9.0-3 is

known to be broken in this regard).

- IBM's xlf compilers: NO known good version that can build/link

the MPI f08 bindings or build/link the OSHMEM Fortran bindings.

- On NetBSD-6 (at least AMD64 and i386), and possibly on OpenBSD,

libtool misidentifies properties of f95/g95, leading to obscure

compile-time failures if used to build Open MPI. You can work

around this issue by ensuring that libtool will not use f95/g95

(e.g., by specifying FC=<some_other_compiler>, or otherwise ensuring

a different Fortran compiler will be found earlier in the path than

f95/g95), or by disabling the Fortran MPI bindings with

--disable-mpi-fortran.

- Absoft 11.5.2 plus a service pack from September 2012 (which Absoft

says is available upon request), or a version later than 11.5.2

(e.g., 11.5.3), is required to compile the new Fortran mpi_f08

module.

- Open MPI does not support the Sparc v8 CPU target. However,

as of Solaris Studio 12.1, and later compilers, one should not

specify -xarch=v8plus or -xarch=v9. The use of the options

-m32 and -m64 for producing 32 and 64 bit targets, respectively,

are now preferred by the Solaris Studio compilers. GCC may

require either "-m32" or "-mcpu=v9 -m32", depending on GCC version.

- It has been noticed that if one uses CXX=sunCC, in which sunCC

is a link in the Solaris Studio compiler release, that the OMPI

build system has issue with sunCC and does not build libmpi_cxx.so.

Therefore the make install fails. So we suggest that one should

use CXX=CC, which works, instead of CXX=sunCC.

- If one tries to build OMPI on Ubuntu with Solaris Studio using the C++

compiler and the -m32 option, you might see a warning:

CC: Warning: failed to detect system linker version, falling back to

custom linker usage

And the build will fail. One can overcome this error by either

setting LD_LIBRARY_PATH to the location of the 32 bit libraries (most

likely /lib32), or giving LDFLAGS="-L/lib32 -R/lib32" to the configure

command. Officially, Solaris Studio is not supported on Ubuntu Linux

distributions, so additional problems might be incurred.

- Open MPI does not support the gccfss compiler (GCC For SPARC

Systems; a now-defunct compiler project from Sun).

- At least some versions of the Intel 8.1 compiler seg fault while

compiling certain Open MPI source code files. As such, it is not

supported.

- The Intel 9.0 v20051201 compiler on IA64 platforms seems to have a

problem with optimizing the ptmalloc2 memory manager component (the

generated code will segv). As such, the ptmalloc2 component will

automatically disable itself if it detects that it is on this

platform/compiler combination. The only effect that this should

have is that the MCA parameter mpi_leave_pinned will be inoperative.

- It has been reported that the Intel 9.1 and 10.0 compilers fail to

compile Open MPI on IA64 platforms. As of 12 Sep 2012, there is

very little (if any) testing performed on IA64 platforms (with any

compiler). Support is "best effort" for these platforms, but it is

doubtful that any effort will be expended to fix the Intel 9.1 /

10.0 compiler issuers on this platform.

- Early versions of the Intel 12.1 Linux compiler suite on x86_64 seem

to have a bug that prevents Open MPI from working. Symptoms

including immediate segv of the wrapper compilers (e.g., mpicc) and

MPI applications. As of 1 Feb 2012, if you upgrade to the latest

version of the Intel 12.1 Linux compiler suite, the problem will go

away.

- Early versions of the Portland Group 6.0 compiler have problems

creating the C++ MPI bindings as a shared library (e.g., v6.0-1).

Tests with later versions show that this has been fixed (e.g.,

v6.0-5).

- The Portland Group compilers prior to version 7.0 require the

"-Msignextend" compiler flag to extend the sign bit when converting

from a shorter to longer integer. This is is different than other

compilers (such as GNU). When compiling Open MPI with the Portland

compiler suite, the following flags should be passed to Open MPI's

configure script:

shell$ ./configure CFLAGS=-Msignextend CXXFLAGS=-Msignextend \

10000 --with-wrapper-cflags=-Msignextend \

--with-wrapper-cxxflags=-Msignextend ...

This will both compile Open MPI with the proper compile flags and

also automatically add "-Msignextend" when the C and C++ MPI wrapper

compilers are used to compile user MPI applications.

- Using the MPI C++ bindings with older versions of the Pathscale

compiler on some platforms is an old issue that seems to be a

problem when Pathscale uses a back-end GCC 3.x compiler. Here's a

proposed solution from the Pathscale support team (from July 2010):

The proposed work-around is to install gcc-4.x on the system and

use the pathCC -gnu4 option. Newer versions of the compiler (4.x

and beyond) should have this fixed, but we'll have to test to

confirm it's actually fixed and working correctly.

We don't anticipate that this will be much of a problem for Open MPI

users these days (our informal testing shows that not many users are

still using GCC 3.x). Contact Pathscale support if you continue to

have problems with Open MPI's C++ bindings.

- Using the Absoft compiler to build the MPI Fortran bindings on Suse

9.3 is known to fail due to a Libtool compatibility issue.

- MPI Fortran API support has been completely overhauled since the

Open MPI v1.5/v1.6 series.

********************************************************************

*** There is now only a single Fortran MPI wrapper compiler and a

*** single Fortran OSHMEM wrapper compiler: mpifort and oshfort,

*** respectively. mpif77 and mpif90 still exist, but they are

*** symbolic links to mpifort.

********************************************************************

*** Similarly, Open MPI's configure script only recognizes the FC

*** and FCFLAGS environment variables (to specify the Fortran

*** compiler and compiler flags, respectively). The F77 and FFLAGS

*** environment variables are IGNORED.

********************************************************************

As a direct result, it is STRONGLY recommended that you specify a

Fortran compiler that uses file suffixes to determine Fortran code

layout (e.g., free form vs. fixed). For example, with some versions

of the IBM XLF compiler, it is preferable to use FC=xlf instead of

FC=xlf90, because xlf will automatically determine the difference

between free form and fixed Fortran source code.

However, many Fortran compilers allow specifying additional

command-line arguments to indicate which Fortran dialect to use.

For example, if FC=xlf90, you may need to use "mpifort --qfixed ..."

to compile fixed format Fortran source files.

You can use either ompi_info or oshmem_info to see with which Fortran

compiler Open MPI was configured and compiled.

There are up to three sets of Fortran MPI bindings that may be

provided depending on your Fortran compiler):

- mpif.h: This is the first MPI Fortran interface that was defined

in MPI-1. It is a file that is included in Fortran source code.

Open MPI's mpif.h does not declare any MPI subroutines; they are

all implicit.

- mpi module: The mpi module file was added in MPI-2. It provides

strong compile-time parameter type checking for MPI subroutines.

- mpi_f08 module: The mpi_f08 module was added in MPI-3. It

provides many advantages over the mpif.h file and mpi module. For

example, MPI handles have distinct types (vs. all being integers).

See the MPI-3 document for more details.

*** The mpi_f08 module is STRONGLY is recommended for all new MPI

Fortran subroutines and applications. Note that the mpi_f08

module can be used in conjunction with the other two Fortran

MPI bindings in the same application (only one binding can be

used per subroutine/function, however). Full interoperability

between mpif.h/mpi module and mpi_f08 module MPI handle types

is provided, allowing mpi_f08 to be used in new subroutines in

legacy MPI applications.

Per the OSHMEM specification, there is only one Fortran OSHMEM binding

provided:

- shmem.fh: All Fortran OpenSHMEM programs **should** include 'shmem.fh',

and Fortran OSHMEM programs that use constants defined by OpenSHMEM

**MUST** include 'shmem.fh'.

The following notes apply to the above-listed Fortran bindings:

- All Fortran compilers support the mpif.h/shmem.fh-based bindings,

with one exception: the MPI_SIZEOF interfaces will only be present

when Open MPI is built with a Fortran compiler that support the

INTERFACE keyword and ISO_FORTRAN_ENV. Most notably, this

excludes the GNU Fortran compiler suite before version 4.9.

- The level of support provided by the mpi module is based on your

Fortran compiler.

If Open MPI is built with a non-GNU Fortran compiler, or if Open

MPI is built with the GNU Fortran compiler >= v4.9, all MPI

subroutines will be prototyped in the mpi module. All calls to

MPI subroutines will therefore have their parameter types checked

at compile time.

If Open MPI is built with an old gfortran (i.e., < v4.9), a

limited "mpi" module will be built. Due to the limitations of

these compilers, and per guidance from the MPI-3 specification,

all MPI subroutines with "choice" buffers are specifically *not*

included in the "mpi" module, and their parameters will not be

checked at compile time. Specifically, all MPI subroutines with

no "choice" buffers are prototyped and will receive strong

parameter type checking at run-time (e.g., MPI_INIT,

MPI_COMM_RANK, etc.).

Similar to the mpif.h interface, MPI_SIZEOF is only supported on

Fortran compilers that support INTERFACE and ISO_FORTRAN_ENV.

- The mpi_f08 module is new and has been tested with the Intel

Fortran compiler and gfortran >= 4.9. Other modern Fortran

compilers may also work (but are, as yet, only lightly tested).

It is expected that this support will mature over time.

Many older Fortran compilers do not provide enough modern Fortran

features to support the mpi_f08 module. For example, gfortran <

v4.9 does provide enough support for the mpi_f08 module.

You can examine the output of the following command to see all

the Fortran features that are/are not enabled in your Open MPI

installation:

shell$ ompi_info | grep -i fort

General Run-Time Support Notes

------------------------------

- The Open MPI installation must be in your PATH on all nodes (and

potentially LD_LIBRARY_PATH (or DYLD_LIBRARY_PATH), if libmpi/libshmem

is a shared library), unless using the --prefix or

--enable-mpirun-prefix-by-default functionality (see below).

- Open MPI's run-time behavior can be customized via MCA ("MPI

Component Architecture") parameters (see below for more information

on how to get/set MCA parameter values). Some MCA parameters can be

set in a way that renders Open MPI inoperable (see notes about MCA

parameters later in this file). In particular, some parameters have

required options that must be included.

- If specified, the "btl" parameter must include the "self"

component, or Open MPI will not be able to deliver messages to the

same rank as the sender. For example: "mpirun --mca btl tcp,self

..."

- If specified, the "btl_tcp_if_exclude" paramater must include the

loopback device ("lo" on many Linux platforms), or Open MPI will

not be able to route MPI messages using the TCP BTL. For example:

"mpirun --mca btl_tcp_if_exclude lo,eth1 ..."

- Running on nodes with different endian and/or different datatype

sizes within a single parallel job is supported in this release.

However, Open MPI does not resize data when datatypes differ in size

(for example, sending a 4 byte MPI_DOUBLE and receiving an 8 byte

MPI_DOUBLE will fail).

MPI Functionality and Features

------------------------------

- Rank reordering support is available using the TreeMatch library. It

is activated for the graph and dist_graph topologies.

- All MPI-3 functionality is supported.

- When using MPI deprecated functions, some compilers will emit

warnings. For example:

shell$ cat deprecated_example.c

#include <mpi.h>

void foo(void) {

MPI_Datatype type;

MPI_Type_struct(1, NULL, NULL, NULL, &type);

}

shell$ mpicc -c deprecated_example.c

deprecated_example.c: In function 'foo':

deprecated_example.c:4: warning: 'MPI_Type_struct' is deprecated (declared at /opt/openmpi/include/mpi.h:1522)

shell$

- MPI_THREAD_MULTIPLE support is included, but is only lightly tested.

It likely does not work for thread-intensive applications. Note

that *only* the MPI point-to-point communication functions for the

BTL's listed here are considered thread safe. Other support

functions (e.g., MPI attributes) have not been certified as safe

when simultaneously used by multiple threads.

- tcp

- sm

- self

Note that Open MPI's thread support is in a fairly early stage; the

above devices may *work*, but the latency is likely to be fairly

high. Specifically, efforts so far have concentrated on

*correctness*, not *performance* (yet).

YMMV.

- MPI_REAL16 and MPI_COMPLEX32 are only supported on platforms where a

portable C datatype can be found that matches the Fortran type

REAL*16, both in size and bit representation.

- The "libompitrace" library is bundled in Open MPI and is installed

by default (it can be disabled via the --disable-libompitrace

flag). This library provides a simplistic tracing of select MPI

function calls via the MPI profiling interface. Linking it in to

your appliation via (e.g., via -lompitrace) will automatically

output to stderr when some MPI functions are invoked:

shell$ cd examples/

shell$ mpicc hello_c.c -o hello_c -lompitrace

shell$ mpirun -np 1 hello_c

MPI_INIT: argc 1

Hello, world, I am 0 of 1

MPI_BARRIER[0]: comm MPI_COMM_WORLD

MPI_FINALIZE[0]

shell$

Keep in mind that the output from the trace library is going to

stderr, so it may output in a slightly different order than the

stdout from your application.

This library is being offered as a "proof of concept" / convenience

from Open MPI. If there is interest, it is trivially easy to extend

it to printf for other MPI functions. Patches and/or suggestions

would be greatfully appreciated on the Open MPI developer's list.

OSHMEM Functionality and Features

------------------------------

- All OpenSHMEM-1.0 functionality is supported.

MPI Collectives

-----------

- The "hierarch" coll component (i.e., an implementation of MPI

collective operations) attempts to discover network layers of

latency in order to segregate individual "local" and "global"

operations as part of the overall collective operation. In this

way, network traffic can be reduced -- or possibly even minimized

(similar to MagPIe). The current "hierarch" component only

separates MPI processes into on- and off-node groups.

Hierarch has had sufficient correctness testing, but has not

received much performance tuning. As such, hierarch is not

activated by default -- it must be enabled manually by setting its

priority level to 100:

mpirun --mca coll_hierarch_priority 100 ...

We would appreciate feedback from the user community about how well

hierarch works for your applications.

- The "fca" coll component: the Mellanox Fabric Collective Accelerator

(FCA) is a solution for offloading collective operations from the

MPI process onto Mellanox QDR InfiniBand switch CPUs and HCAs.

- The "ML" coll component is an implementation of MPI collective

operations that takes advantage of communication hierarchies in

modern systems. A ML collective operation is implemented by

combining multiple independently progressing collective primitives

implemented over different communication hierarchies, hence a ML

collective operation is also referred to as a hierarchical

collective operation. The number of collective primitives that are

included in a ML collective operation is a function of

subgroups(hierarchies). Typically, MPI processes in a single

communication hierarchy such as CPU socket, node, or subnet are

grouped together into a single subgroup (hierarchy). The number of

subgroups are configurable at runtime, and each different collective

operation could be configured to have a different of number of

subgroups.

The component frameworks and components used by/required for a

"ML" collective operation.

Frameworks:

* "sbgp" - Provides functionality for grouping processes into

subgroups

* "bcol" - Provides collective primitives optimized for a particular

communication hierarchy

Components:

* sbgp components - Provides grouping functionality over a CPU

socket ("basesocket"), shared memory

("basesmuma"), Mellanox's ConnectX HCA

("ibnet"), and other interconnects supported by

PML ("p2p")

* BCOL components - Provides optimized collective primitives for

shared memory ("basesmuma"), Mellanox's ConnectX

HCA ("iboffload"), and other interconnects

supported by PML ("ptpcoll")

- The "cuda" coll component provides CUDA-aware support for the

reduction type collectives with GPU buffers. This component is only

compiled into the library when the library has been configured with

CUDA-aware support. It intercepts calls to the reduction

collectives, copies the data to staging buffers if GPU buffers, then

calls underlying collectives to do the work.

OSHMEM Collectives

-----------

- The "fca" scoll component: the Mellanox Fabric Collective Accelerator

(FCA) is a solution for offloading collective operations from the

MPI process onto Mellanox QDR InfiniBand switch CPUs and HCAs.

- The "basic" scoll component: Reference implementation of all OSHMEM

collective operations.

Network Support

---------------

- There are three main MPI network models available: "ob1", "cm", and

"yalla". "ob1" uses BTL ("Byte Transfer Layer") components for each

supported network. "cm" uses MTL ("Matching Tranport Layer")

components for each supported network. "yalla" uses the Mellanox

MXM transport.

- "ob1" supports a variety of networks that can be used in

combination with each other (per OS constraints; e.g., there are

reports that the GM and OpenFabrics kernel drivers do not operate

well together):

- OpenFabrics: InfiniBand, iWARP, and RoCE

- Loopback (send-to-self)

- Shared memory

- TCP

- Intel Phi SCIF

- SMCUDA

- Cisco usNIC

- uGNI (Cray Gemini, Aries)

- vader (XPMEM, Linux CMA, Linux KNEM, and general shared memory)

- "cm" supports a smaller number of networks (and they cannot be

used together), but may provide better overall MPI performance:

- QLogic InfiniPath / Intel True Scale PSM

- Intel Omni-Path PSM2

- Mellanox MXM

- Portals4

- OpenFabrics Interfaces ("libfabric" tag matching)

Open MPI will, by default, choose to use "cm" when one of the

above transports can be used. Otherwise, "ob1" will be used and

the corresponding BTLs will be selected. Users can force the use

of ob1 or cm if desired by setting the "pml" MCA parameter at

run-time:

shell$ mpirun --mca pml ob1 ...

shell$ mpirun --mca pml cm ...

- Similarly, there are two OSHMEM network models available: "yoda",

and "ikrit". "yoda" also uses the BTL components for many supported

network. "ikrit" interfaces directly with Mellanox MXM.

- "yoda" supports a variety of networks that can be used:

- OpenFabrics: InfiniBand, iWARP, and RoCE

- Loopback (send-to-self)

- Shared memory

- TCP

- "ikrit" only supports Mellanox MXM.

- MXM is the Mellanox Messaging Accelerator library utilizing a full

range of IB transports to provide the following messaging services

to the upper level MPI/OSHMEM libraries:

- Usage of all available IB transports

- Native RDMA support

- Progress thread

- Shared memory communication

- Hardware-assisted reliability

- The usnic BTL is support for Cisco's usNIC device ("userspace NIC")

on Cisco UCS servers with the Virtualized Interface Card (VIC).

Although the usNIC is accessed via the OpenFabrics Libfabric API

stack, this BTL is specific to the Cisco usNIC device.

- uGNI is a Cray library for communicating over the Gemini and Aries

interconnects.

- The OpenFabrics Enterprise Distribution (OFED) software package v1.0

will not work properly with Open MPI v1.2 (and later) due to how its

Mellanox InfiniBand plugin driver is created. The problem is fixed

OFED v1.1 (and later).

- Better memory management support is available for OFED-based

transports using the "ummunotify" Linux kernel module. OFED memory

managers are necessary for better bandwidth when re-using the same

buffers for large messages (e.g., benchmarks and some applications).

Unfortunately, the ummunotify module was not accepted by the Linux

kernel community (and is still not distributed by OFED). But it

still remains the best memory management solution for MPI

applications that used the OFED network transports. If Open MPI is

able to find the <linux/ummunotify.h> header file, it will build

support for ummunotify and include it by default. If MPI processes

then find the ummunotify kernel module loaded and active, then their

memory managers (which have been shown to be problematic in some

cases) will be disabled and ummunotify will be used. Otherwise, the

same memory managers from prior versions of Open MPI will be used.

The ummunotify Linux kernel module can be downloaded from:

http://lwn.net/Articles/343351/

- The use of fork() with OpenFabrics-based networks (i.e., the openib

BTL) is only partially supported, and only on Linux kernels >=

v2.6.15 with libibverbs v1.1 or later (first released as part of

OFED v1.2), per restrictions imposed by the OFED network stack.

- Linux "knem" support is used when the "vader" or "sm" (shared

memory) BTLs are compiled with knem support (see the --with-knem

configure option) and the knem Linux module is loaded in the running

kernel. If the knem Linux kernel module is not loaded, the knem

support is (by default) silently deactivated during Open MPI jobs.

See http://runtime.bordeaux.inria.fr/knem/ for details on Knem.

- Linux Cross-Memory Attach (CMA) or XPMEM is used by the vader

shared-memory BTL when the CMA/XPMEM libraries are installedm,

respectively. Linux CMA and XPMEM are similar (but different)

mechanisms for Open MPI to utilize single-copy semantics for shared

memory.

Open MPI Extensions

-------------------

- An MPI "extensions" framework has been added (but is not enabled by

default). See the "Open MPI API Extensions" section below for more

information on compiling and using MPI extensions.

- The following extensions are included in this version of Open MPI:

- affinity: Provides the OMPI_Affinity_str() routine on retrieving

a string that contains what resources a process is bound to. See

its man page for more details.

- cr: Provides routines to access to checkpoint restart routines.

See ompi/mpiext/cr/mpiext_cr_c.h for a listing of availble

functions.

- cuda: When the library is compiled with CUDA-aware support, it provides

two things. First, a macro MPIX_CUDA_AWARE_SUPPORT. Secondly, the

function MPIX_Query_cuda_support that can be used to query for support.

- example: A non-functional extension; its only purpose is to

provide an example for how to create other extensions.

===========================================================================

Building Open MPI

-----------------

Open MPI uses a traditional configure script paired with "make" to

build. Typical installs can be of the pattern:

---------------------------------------------------------------------------

shell$ ./configure [...options...]

shell$ make all install

---------------------------------------------------------------------------

There are many available configure options (see "./configure --help"

for a full list); a summary of the more commonly used ones is included

below.

Note that for many of Open MPI's --with-<foo> options, Open MPI will,

by default, search for header files and/or libraries for <foo>. If

the relevant files are found, Open MPI will built support for <foo>;

if they are not found, Open MPI will skip building support for <foo>.

However, if you specify --with-<foo> on the configure command line and

Open MPI is unable to find relevant support for <foo>, configure will

assume that it was unable to provide a feature that was specifically

requested and will abort so that a human can resolve out the issue.

INSTALLATION OPTIONS

--prefix=<directory>

Install Open MPI into the base directory named <directory>. Hence,

Open MPI will place its executables in <directory>/bin, its header

files in <directory>/include, its libraries in <directory>/lib, etc.

--disable-shared

By default, libmpi and libshmem are built as a shared library, and

all components are built as dynamic shared objects (DSOs). This

switch disables this default; it is really only useful when used with

--enable-static. Specifically, this option does *not* imply

--enable-static; enabling static libraries and disabling shared

libraries are two independent options.

--enable-static

Build libmpi and libshmem as static libraries, and statically link in all

components. Note that this option does *not* imply

--disable-shared; enabling static libraries and disabling shared

libraries are two independent options.

Be sure to read the description of --without-memory-manager, below;

it may have some effect on --enable-static.

--disable-wrapper-rpath

By default, the wrapper compilers (e.g., mpicc) will enable "rpath"

support in generated executables on systems that support it. That

is, they will include a file reference to the location of Open MPI's

libraries in the application executable itself. This means that

the user does not have to set LD_LIBRARY_PATH to find Open MPI's

libraries (e.g., if they are installed in a location that the

run-time linker does not search by default).

On systems that utilize the GNU ld linker, recent enough versions

will actually utilize "runpath" functionality, not "rpath". There

is an important difference between the two:

"rpath": the location of the Open MPI libraries is hard-coded into

the MPI/OSHMEM application and cannot be overridden at run-time.

"runpath": the location of the Open MPI libraries is hard-coded into

the MPI/OSHMEM application, but can be overridden at run-time by

setting the LD_LIBRARY_PATH environment variable.

For example, consider that you install Open MPI vA.B.0 and

compile/link your MPI/OSHMEM application against it. Later, you install

Open MPI vA.B.1 to a different installation prefix (e.g.,

/opt/openmpi/A.B.1 vs. /opt/openmpi/A.B.0), and you leave the old

installation intact.

In the rpath case, your MPI application will always use the

libraries from your A.B.0 installation. In the runpath case, you

can set the LD_LIBRARY_PATH environment variable to point to the

A.B.1 installation, and then your MPI application will use those

libraries.

Note that in both cases, however, if you remove the original A.B.0

installation and set LD_LIBRARY_PATH to point to the A.B.1

installation, your application will use the A.B.1 libraries.

This rpath/runpath behavior can be disabled via

--disable-wrapper-rpath.

--enable-dlopen

Build all of Open MPI's components as standalone Dynamic Shared

Objects (DSO's) that are loaded at run-time (this is the default).

The opposite of this option, --disable-dlopen, causes two things:

1. All of Open MPI's components will be built as part of Open MPI's

normal libraries (e.g., libmpi).

2. Open MPI will not attempt to open any DSO's at run-time.

Note that this option does *not* imply that OMPI's libraries will be

built as static objects (e.g., libmpi.a). It only specifies the

location of OMPI's components: standalone DSOs or folded into the

Open MPI libraries. You can control whether Open MPI's libraries

are build as static or dynamic via --enable|disable-static and

--enable|disable-shared.

--with-platform=FILE

Load configure options for the build from FILE. Options on the

command line that are not in FILE are also used. Options on the

command line and in FILE are replaced by what is in FILE.

NETWORKING SUPPORT / OPTIONS

--with-fca=<directory>

Specify the directory where the Mellanox FCA library and

header files are located.

FCA is the support library for Mellanox QDR switches and HCAs.

--with-hcoll=<directory>

Specify the directory where the Mellanox hcoll library and header

files are located. This option is generally only necessary if the

hcoll headers and libraries are not in default compiler/linker

search paths.

hcoll is the support library for MPI collective operation offload on

Mellanox ConnectX-3 HCAs (and later).

--with-knem=<directory>

Specify the directory where the knem libraries and header files are

located. This option is generally only necessary if the knem headers

and libraries are not in default compiler/linker search paths.

knem is a Linux kernel module that allows direct process-to-process

memory copies (optionally using hardware offload), potentially

increasing bandwidth for large messages sent between messages on the

same server. See http://runtime.bordeaux.inria.fr/knem/ for

details.

--with-libfabric=<directory>

Specify the directory where the OpenFabrics Interfaces libfabric

library and header files are located. This option is generally only

necessary if the libfabric headers and libraries are not in default

compiler/linker search paths.

Libfabric is the support library for OpenFabrics Interfaces-based

network adapters, such as Cisco usNIC, Intel True Scale PSM, etc.

--with-libfabric-libdir=<directory>

Look in directory for the libfabric libraries. By default, Open MPI

will look in <libfabric directory>/lib and <libfabric

directory>/lib64, which covers most cases. This option is only

needed for special configurations.

--with-mxm=<directory>

Specify the directory where the Mellanox MXM library and header

files are located. This option is generally only necessary if the

MXM headers and libraries are not in default compiler/linker search

paths.

MXM is the support library for Mellanox Network adapters.

--with-mxm-libdir=<directory>

Look in directory for the MXM libraries. By default, Open MPI will

look in <mxm directory>/lib and <mxm directory>/lib64, which covers

most cases. This option is only needed for special configurations.

--with-portals4=<directory>

Specify the directory where the Portals4 libraries and header files

are located. This option is generally only necessary if the Portals4

headers and libraries are not in default compiler/linker search

paths.

Portals is a low-level network API for high-performance networking

on high-performance computing systems developed by Sandia National

Laboratories, Intel Corporation, and the University of New Mexico.

The Portals 4 Reference Implementation is a complete implementation

of Portals 4, with transport over InfiniBand verbs and UDP.

--with-portals4-libdir=<directory>

Location of libraries to link with for Portals4 support.

--with-portals4-max-md-size=SIZE

--with-portals4-max-va-size=SIZE

Set configuration values for Portals 4

--with-psm=<directory>

Specify the directory where the QLogic InfiniPath / Intel True Scale

PSM library and header files are located. This option is generally

only necessary if the PSM headers and libraries are not in default

compiler/linker search paths.

PSM is the support library for QLogic InfiniPath and Intel TrueScale

network adapters.

--with-psm-libdir=<directory>

Look in directory for the PSM libraries. By default, Open MPI will

look in <psm directory>/lib and <psm directory>/lib64, which covers

most cases. This option is only needed for special configurations.

--with-psm2=<directory>

Specify the directory where the Intel Omni-Path PSM2 library and

header files are located. This option is generally only necessary

if the PSM2 headers and libraries are not in default compiler/linker

search paths.

PSM2 is the support library for Intel Omni-Path network adapters.

--with-psm2-libdir=<directory>

Look in directory for the PSM2 libraries. By default, Open MPI will

look in <psm2 directory>/lib and <psm2 directory>/lib64, which

covers most cases. This option is only needed for special

configurations.

--with-scif=<dir>

Look in directory for Intel SCIF support libraries

--with-verbs=<directory>

Specify the directory where the verbs (also know as OpenFabrics, and

previously known as OpenIB) libraries and header files are located.

This option is generally only necessary if the verbs headers and

libraries are not in default compiler/linker search paths.

"OpenFabrics" refers to operating system bypass networks, such as

InfiniBand, usNIC, iWARP, and RoCE (aka "IBoIP").

--with-verbs-libdir=<directory>

Look in directory for the verbs libraries. By default, Open MPI

will look in <verbs_directory>/lib and <verbs_ directory>/lib64,

which covers most cases. This option is only needed for special

configurations.

--with-verbs-usnic

This option will activate support in Open MPI for disabling a

dire-sounding warning message from libibverbs that Cisco usNIC

devices are not supported (because Cisco usNIC devices are supported

through libfabric, not libibverbs). This libibverbs warning can

also be suppressed by installing the "no op" libusnic_verbs plugin

for libibverbs (see https://github.com/cisco/libusnic_verbs, or

download binaries from cisco.com). This option is disabled by

default because it causes libopen-pal.so to depend on libibverbs.so,

which is undesirable to many downstream packagers.

--with-usnic

Abort configure if Cisco usNIC support cannot be built.

RUN-TIME SYSTEM SUPPORT

--enable-mpirun-prefix-by-default

This option forces the "mpirun" command to always behave as if

"--prefix $prefix" was present on the command line (where $prefix is

the value given to the --prefix option to configure). This prevents

most rsh/ssh-based users from needing to modify their shell startup

files to set the PATH and/or LD_LIBRARY_PATH for Open MPI on remote

nodes. Note, however, that such users may still desire to set PATH

-- perhaps even in their shell startup files -- so that executables

such as mpicc and mpirun can be found without needing to type long

path names. --enable-orterun-prefix-by-default is a synonym for

this option.

--enable-sensors

Enable internal sensors (default: disabled).

--enable-orte-static-ports

Enable orte static ports for tcp oob (default: enabled).

--with-alps

Force the building of for the Cray Alps run-time environment. If

Alps support cannot be found, configure will abort.

--with-loadleveler

Force the building of LoadLeveler scheduler support. If LoadLeveler

support cannot be found, configure will abort.

--with-lsf=<directory>

Specify the directory where the LSF libraries and header files are

located. This option is generally only necessary if the LSF headers

View remainder of file in raw view

FilesExpand file tree

README

Latest commit

History

README

File metadata and controls