CHANGELOG.md 16.8 KB
Newer Older
Gilbert Lee's avatar
Gilbert Lee committed
1
2
# Changelog for TransferBench

3
Documentation for TransferBench is available at
Lisa's avatar
Lisa committed
4
[https://rocm.docs.amd.com/projects/TransferBench](https://rocm.docs.amd.com/projects/TransferBench).
Lisa Delaney's avatar
Lisa Delaney committed
5

gilbertlee-amd's avatar
gilbertlee-amd committed
6
7
8
9
10
11
12
13
## v1.52
### Added
- Added USE_HSA_DMA env var to switch to using hsa_amd_memory_async_copy instead of hipMemcpyAsync for DMA execution
- Added ability to set USE_GPU_DMA env var for a2a benchmark
- Adding check for large BAR enablement for GPU devices during topology check
### Fixed
- Potential memory leak if HSA reports 0 hops between GPUs and CPUs

14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
## v1.51

## Modified
- CSV output has been modified slightly to match normal terminal output
- Output for non single stream mode has been changed to match single stream mode (results per Executor)

### Added
- Support for sub-iterations via NUM_SUBITERATIONS.  This allows for additional looping during an iteration
  If set to 0, this should infinitely loop (which may be useful for some debug purposes)
- Support for variable number of subexecutors (currently for GPU-GFX executor only).  Setting subExecutors to
  0 will run over a range of CUs to use, and report only the results of the best one found. This can be tuned
  for performance by setting the MIN_VAR_SUBEXEC and MAX_VAR_SUBEXEC environment variables to narrow the
  search space.  The number of CUs used will be identical for all variable subExecutor transfers
- Experimental new "healthcheck" preset config which currently only supports MI300 series.  This preset runs
  through CPU to GPU bandwidth tests and all-to-all XGMI bandwidth tests and compares against expected values
  Pass criteria limits can be modified (due to platform differences) via the environment variables
  LIMIT_UDIR (undirectional), LIMIT_BDIR (bidirectional), and LIMIT_A2A (Per GPU-GPU link bandwidth)

### Fixed
- Fixed out-of-bounds memory access during topology detection that can happen if the number of
  CPUs is less than the number of NUMA domains
- Fixed CU masking functionality on multi-XCD architectures (e.g. MI300)

37
38
39
40
41
42
43
44
45
## v1.50

### Added
- Adding new parallel copy preset benchmark (pcopy)
  - Usage: ./TransferBench pcopy <numBytes=64M> <#CUs=8> <srcGpu=0> <minGpus=1> <maxGpus=#GPU-1>
### Fixed
- Removed non-copies DMA Transfers (this had previously been using hipMemset)
- Fixed CPU executor when operating on null destination

46
47
48
49
50
## v1.49

### Fixes
* Enumerating previously missed DMA engines used only for CPU traffic in topology display

51
52
53
54
55
56
57
58
59
60
61
62
## v1.48

### Fixes
* Various fixes for TransferBenchCuda

### Additions
* Support for targeting specific DMA engines via executor subindex (e.g. D0.1)
* Printing warnings when exeuctors are overcommited

### Modifications
* USE_REMOTE_READ supported for rwrite preset benchmark

63
64
65
66
67
## v1.47

### Fixes
* Fixing CUDA support

68
69
70
71
72
73
74
75
76
## v1.46

### Fixes
* Fixing GFX_UNROLL set to 13 (past 8) on gfx906 cards

### Modifications
* GFX_SINGLE_TEAM=1 by default
* Adding field showing summation of individual Transfer bandwidths for Executors

gilbertlee-amd's avatar
gilbertlee-amd committed
77
78
79
80
81
82
83
84
85
86
## v1.45

### Additions
* Adding A2A_MODE to a2a preset (0 = copy, 1 = read-only, 2 = write-only)
* Adding GFX_UNROLL to modify GFX kernel's unroll factor
* Adding GFX_WAVE_ORDER to modify order in which wavefronts process data

### Modifications
* Rewrote the GFX reduction kernel to support new wave ordering

87
88
89
90
91
92
## v1.44

### Additions
* Adding rwrite preset to benchmark remote parallel writes
 * Usage: ./TransferBench rwrite <numBytes=64M> <#CUs=8> <srcGpu=0> <minGpus=1> <maxGpus=3>

93
94
95
96
97
## v1.43

### Changes
* Modifying a2a to show executor timing, as well as executor min/max bandwidth

98
99
100
101
102
103
## v1.42

### Fixes
* Fixing schmoo maxNumCus optional arg parsing
* Schmoo output modified to be easier to copy

104
105
106
107
108
109
110
111
112
## v1.41

### Additions
* Adding schmoo preset config benchmarks local/remote reads/writes/copies
  * Usage: ./TransferBench schmoo <numBytes=64M> <localIdx=0> <remoteIdx=1> <maxNumCUs=32>

### Fixes
* Fixing some misreported timings when running with non-fixed number of iterations

113
114
115
116
117
## v1.40

### Fixes
* Fixing XCC defaulting to 0 instead of random for preset configs, ignoring XCC_PREF_TABLE

118
119
120
121
122
123
124
## v1.39

### Additions
* (Experimental) Adding support for Executor sub-index
### Fixes
- Remove deprecated gcnArch code.  ROCm version must include support for hipDeviceMallocUncached

125
126
127
128
129
## v1.38

### Fixes
* Adding missing threadfence which could cause non-fine-grained Transfers to report higher speeds

130
131
132
133
134
135
136
137
## v1.37

### Changes
* USE_SINGLE_STREAM is enabled by default now.  (Disable via USE_SINGLE_STREAM=0)

### Fixes
* Fix unrecognized token error when XCC_PREF_TABLE is unspecified

138
139
140
141
142
143
144
## v1.36

### Additions

* (Experimental) Adding XCC filtering - combined with XCC_PREF_TABLE, this tries to select
  specific XCCs to use for specific (SRC->DST) Transfers

145
146
147
148
149
## v1.35

### Additions

* USE_FINE_GRAIN also applies to a2a preset
150

151
## v1.34
152
153
154
155

### Additions

* Set `GPU_KERNEL=3` as default for gfx942
156

157
## v1.33
Lisa Delaney's avatar
Lisa Delaney committed
158
159
160
161
162

### Additions

* Added the `ALWAYS_VALIDATE` environment variable to allow for validation after every iteration, instead
  of only once at the end of all iterations
163

164
## v1.32
Lisa Delaney's avatar
Lisa Delaney committed
165
166
167
168

### Changes

* Increased the line limit from 2048 to 32768
169

gilbertlee-amd's avatar
gilbertlee-amd committed
170
## v1.31
Lisa Delaney's avatar
Lisa Delaney committed
171
172
173
174
175

### Changes

* `SHOW_ITERATIONS` now shows XCC:CU instead of just CU ID
* `SHOW_ITERATIONS` is printed when `USE_SINGLE_STREAM`=1
gilbertlee-amd's avatar
gilbertlee-amd committed
176

177
## v1.30
Lisa Delaney's avatar
Lisa Delaney committed
178
179
180
181
182
183
184
185
186

### Additions

* `BLOCK_SIZE` has been added to control the threadblock size (must be a multiple of 64, up to 512)
* `BLOCK_ORDER` has been added to control how work is ordered for GFX-executors running
  `USE_SINGLE_STREAM`=1
  * 0 - Threadblocks for transfers are ordered sequentially (default)
  * 1 - Threadblocks for transfers are interleaved
  * 2 - Threadblocks for transfers are ordered randomly
187

188
## v1.29
Lisa Delaney's avatar
Lisa Delaney committed
189
190
191
192
193
194
195
196
197
198
199
200
201
202

### Additions

* A2A preset config now responds to `USE_REMOTE_READ`

### Fixes

* Race-condition during wall-clock initialization caused "inf" during single-stream runs
* CU numbering output after CU masking

### Changes

* The default number of warmups has been reverted to 3
* The default unroll factor for gfx940/941 has been set to 6
203

204
## v1.28
Lisa Delaney's avatar
Lisa Delaney committed
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221

### Additions

* Added `A2A_DIRECT`, which only runs all-to-all on directly connected GPUs (now on by default)
* Added average statistics for P2P and A2A benchmarks
* Added `USE_FINE_GRAIN` for P2P benchmark
  * With older devices, P2P performance with default coarse-grain device memory stops timing as soon
    as a request is sent to data fabric, and not actually when it arrives remotely. This can artificially
    inflate bandwidth numbers, especially when sending small amounts of data.

### Changes

* Modified P2P output to help distinguish between CPU and GPU devices

### Fixes

* Fixed Makefile target to prevent unnecessary re-compilation
222

223
## v1.27
Lisa Delaney's avatar
Lisa Delaney committed
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242

### Additions

* Added cmdline preset to allow specification of  simple tests on command line (e.g.,
  `./TransferBench cmdline 64M "1 4 G0->G0->G1"`)
* Adding the `HIDE_ENV` environment variable, which stops environment variable values from printing
* Adding the `CU_MASK` environment variable, which allows you to select the CUs to run on
* `CU_MASK` is specified in CU indices (0-#CUs-1), where ' - ' can be used to denote ranges of values
  (e.g., `CU_MASK`=3-8,16 requests that transfer be run only on CUs 3,4,5,6,7,8,16)
  * Note that this is somewhat experimental and may not work on all hardware
* `SHOW_ITERATIONS` now shows CU usage for that iteration (experimental)

### Changes

* Added extra comments on commonly missing includes with details on how to install them

### Fixes

* CUDA compilation works again (the `wall_clock64` CUDA alias was not defined)
243

244
245
## v1.26

Lisa Delaney's avatar
Lisa Delaney committed
246
247
248
249
250
251
252
253
254
255
256
257
### Additions

* Setting SHOW_ITERATIONS=1 provides additional information about per-iteration timing for file and
  P2P configs
  * For file configs, iterations are sorted from min to max bandwidth and displayed with standard
    deviation
  * For P2P, min/max/standard deviation is shown for each direction

### Changes

* P2P benchmark formatting now reports bidirectional bandwidth in each direction (as well as sum) for
  clarity
258

259
## v1.25
Lisa Delaney's avatar
Lisa Delaney committed
260
261
262
263
264

### Fixes

* Fixed a bug in the P2P bidirectional benchmark that used the incorrect number of `subExecutors` for
  CPU<->GPU tests
265

gilbertlee-amd's avatar
gilbertlee-amd committed
266
## v1.24
Lisa Delaney's avatar
Lisa Delaney committed
267
268
269
270
271

### Additions

* New All-To-All GPU benchmark accessed by preset "A2A"
* Added gfx941 wall clock frequency
gilbertlee-amd's avatar
gilbertlee-amd committed
272

273
## v1.23
Lisa Delaney's avatar
Lisa Delaney committed
274
275
276
277
278

### Additions

* New GPU subexec scaling benchmark accessed by preset "scaling"
  * Tests GPU-GFX copy performance based on # of CUs used
279

280
## v1.22
Lisa Delaney's avatar
Lisa Delaney committed
281
282
283
284

### Changes

* Switched the kernel timing function to `wall_clock64`
285

gilbertlee-amd's avatar
gilbertlee-amd committed
286
## v1.21
Lisa Delaney's avatar
Lisa Delaney committed
287
288
289
290

### Fixes

* Fixed a bug with `SAMPLING_FACTOR`
gilbertlee-amd's avatar
gilbertlee-amd committed
291

292
## v1.20
Lisa Delaney's avatar
Lisa Delaney committed
293
294
295
296
297

### Fixes

* `VALIDATE_DIRECT` can now be used with `USE_PREP_KERNEL`
* Switched to local GPU for validating GPU memory
298

gilbertlee-amd's avatar
gilbertlee-amd committed
299
## v1.19
Lisa Delaney's avatar
Lisa Delaney committed
300
301
302
303
304

### Additions

* `VALIDATE_DIRECT` now also applies to source memory array checking
* Added null memory pointer check prior to deallocation
gilbertlee-amd's avatar
gilbertlee-amd committed
305

306
## v1.18
Lisa Delaney's avatar
Lisa Delaney committed
307
308
309
310
311
312
313
314
315
316
317
318
319

### Additions

* Adding the ability to validate GPU destination memory directly without going through the CPU
  staging buffer (`VALIDATE_DIRECT`)
  * Note that this only works on AMD devices with large-bar access enabled, and may slow things down
    considerably

### Changes

* Refactored how environment variables are displayed
* Mismatch stops after the first detected error within an array instead of listing all mismatched
  elements
320

321
## v1.17
Lisa Delaney's avatar
Lisa Delaney committed
322
323
324
325
326
327
328
329
330
331
332
333
334
335

### Additions

* Allowed switch to GFX kernel for source array initialization (`USE_PREP_KERNEL`)
  * Note that `USE_PREP_KERNEL` can't be used with `FILL_PATTERN`
* Added the ability to compile with nvcc only (`TransferBenchCuda`)

### Changes

* The default pattern was set to [Element i = ((i * 517) modulo 383 + 31) * (srcBufferIdx + 1)]

### Fixes

* Added the `example.cfg` file
336

337
## v1.16
Lisa Delaney's avatar
Lisa Delaney committed
338
339
340
341
342
343
344

### Additions

* Additional src array validation during preparation
* Added a new environment variable (`CONTINUE_ON_ERROR`) to resume tests after a mis-match
  detection
* Initialized GPU memory to 0 during allocation
345

346
## v1.15
Lisa Delaney's avatar
Lisa Delaney committed
347
348
349
350
351
352
353
354
355

### Fixes

* Fixed a bug that prevented single transfers greater than 8 GB

### Changes

* Removed "check for latest ROCm" warning when allocating too much memory
* Off-source memory value is now printed when a mis-match is detected
356

PedramAlizadeh's avatar
PedramAlizadeh committed
357
## v1.14
Lisa Delaney's avatar
Lisa Delaney committed
358
359
360
361
362
363

### Additions

* Added documentation
* Added pthread linking in src/Makefile and CMakeLists.txt
* Added printing off the hex value of the floats for output and reference
364

PedramAlizadeh's avatar
PedramAlizadeh committed
365
366
## v1.13

Lisa Delaney's avatar
Lisa Delaney committed
367
368
369
370
371
372
373
### Additions

* Added support for cmake

### Changes

* Converted to the Pitchfork layout standard
PedramAlizadeh's avatar
PedramAlizadeh committed
374

375
## v1.12
Lisa Delaney's avatar
Lisa Delaney committed
376
377
378
379
380

### Additions

* Added support for TransferBench on NVIDIA platforms (via `HIP_PLATFORM`=nvidia)
  * Note that CPU executors on NVIDIA platform cannot access GPU memory (no large-bar access)
381

gilbertlee-amd's avatar
gilbertlee-amd committed
382
## v1.11
Lisa Delaney's avatar
Lisa Delaney committed
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

### Additions

* Added multi-input/multi-output (MIMO) support: transfers now can reduce (element-wise summation)
  multiple input memory arrays and write sums to multiple outputs
* Added GPU-DMA executor 'D', which uses `hipMemcpy` for SDMA copies
  * Previously, this was done using `USE_HIP_CALL`, but now GPU-GFX kernel can run in parallel with
    GPU-DMA, instead of applying to all GPU executors globally
  * GPU-DMA executor can only be used for single-input/single-output transfers
  * GPU-DMA executor can only be associated with one SubExecutor
* Added new "Null" memory type 'N', which represents empty memory. This allows for read-only or
  write-only transfers
* Added new `GPU_KERNEL` environment variable that allows switching between various GPU-GFX
  reduction kernels

### Optimizations

* Improved GPU-GFX kernel performance based on hardware architecture when running with
  fewer CUs

### Changes

* Updated the `example.cfg` file to cover new features
* Updated output to support MIMO
* Changed CU and CPU thread naming to SubExecutors for consistency
* Sweep Preset: default sweep preset executors now includes DMA
* P2P benchmarks:
  * Removed `p2p_rr`, `g2g` and `g2g_rr` (now only works via P2P)
    * Setting `NUM_CPU_DEVICES`=0 can only be used to benchmark GPU devices (like `g2g`)
    * The new `USE_REMOTE_READ` environment variable replaces `_rr` presets
  * New environment variable `USE_GPU_DMA`=1 replaces `USE_HIP_CALL`=1 for benchmarking with
    GPU-DMA Executor
  * Number of GPU SubExecutors for benchmark can be specified via `NUM_GPU_SE`
    * Defaults to all CUs for GPU-GFX, 1 for GPU-DMA
  * Number of CPU SubExecutors for benchmark can be specified via `NUM_CPU_SE`
* Psuedo-random input pattern has been slightly adjusted to have different patterns for each input
  array within same transfer

### Removals

* `USE_HIP_CALL`: use `GPU-DMA` executor 'D' or set `USE_GPU_DMA`=1 for P2P
  benchmark presets
  * Currently, a warning will be issued if `USE_HIP_CALL` is set to 1 and the program will stop
* `NUM_CPU_PER_TRANSFER`: the number of CPU SubExecutors will be whatever is specified for the
  transfer
* `USE_MEMSET`: this function can now be done via a transfer using the null memory type
gilbertlee-amd's avatar
gilbertlee-amd committed
429

430
## v1.10
Lisa Delaney's avatar
Lisa Delaney committed
431
432
433
434

### Fixes

* Fixed incorrect bandwidth calculation when using single stream mode and per-transfer data sizes
435

436
## v1.09
Lisa Delaney's avatar
Lisa Delaney committed
437
438
439
440
441
442
443
444

### Additions

* Printing off src/dst memory addresses during interactive mode

### Changes

* Switching to `numa_set_preferred` instead of `set_mempolicy`
445

446
## v1.08
Lisa Delaney's avatar
Lisa Delaney committed
447
448
449
450
451
452

### Changes

* Fixed handling of non-configured NUMA nodes
* Topology detection now shows actual NUMA node indices
* Fixed 'for' issue with `NUM_GPU_DEVICES`
453

454
## v1.07
Lisa Delaney's avatar
Lisa Delaney committed
455
456
457
458

### Fixes

* Fixed bug with allocations involving non-default CPU memory types
459

gilbertlee-amd's avatar
gilbertlee-amd committed
460
## v1.06
Lisa Delaney's avatar
Lisa Delaney committed
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476

### Additions

* Unpinned CPU memory type ('U'), which may require `HSA_XNACK`=1 in order to access via
  GPU executors
* Added sweep configuration logging to `lastSweep.cfg`
* Ability to specify the number of CUs to use for sweep-based presets

### Changes

* Modified advanced configuration file format to accept bytes-per-transfer

### Fixes

* Fixed random sweep repeatability
* Fixed bug with CPU NUMA node memory allocation
gilbertlee-amd's avatar
gilbertlee-amd committed
477

478
## v1.05
Lisa Delaney's avatar
Lisa Delaney committed
479
480
481
482
483
484
485
486
487

### Additions

* Topology output now includes NUMA node information
* Support for NUMA nodes with no CPU cores (e.g., CXL memory)

### Removals

* The `SWEEP_SRC_IS_EXE` environment variable was removed
488

489
## v1.04
Lisa Delaney's avatar
Lisa Delaney committed
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504

### Additions

* There are new environment variables for sweep based presets:
  * `SWEEP_XGMI_MIN`: The minumum number of XGMI hops for transfers
  * `SWEEP_XGMI_MAX`: The maximum number of XGMI hops for transfers
  * `SWEEP_SEED`: Uses a random seed
  * `SWEEP_RAND_BYTES`: Uses a random amount of bytes (up to pre-specified N) for each transfer

### Changes

* CSV output for sweep now includes an environment variables section followed by output
* CSV output no longer lists environment variable parameters in columns
* We changed the default number of warmup iterations from 3 to 1
* Split CSV output of link type to `ExeToSrcLinkType` and `ExeToDstLinkType`
505

Gilbert Lee's avatar
Gilbert Lee committed
506
## v1.03
Lisa Delaney's avatar
Lisa Delaney committed
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529

### Additions

* There are new preset modes stress-test benchmarks: `sweep` and `randomsweep`
  * `sweep` iterates over all possible sets of transfers to test
  * `randomsweep` iterates over random sets of transfers
  * New sweep-only environment variables can modify `sweep`
    * `SWEEP_SRC`: String containing only "B","C","F", or "G" that defines possible source memory types
    * `SWEEP_EXE`: String containing only "C" or "G" that defines possible executors
    * `SWEEP_DST`: String containing only "B","C","F", or "G" that defines possible destination memory types
    * `SWEEP_SRC_IS_EXE`: Restrict the executor to be the same as the source, if non-zero
    * `SWEEP_MIN`: Minimum number of parallel transfers to test
    * `SWEEP_MAX`: Maximum number of parallel transfers to test
    * `SWEEP_COUNT`: Maximum number of tests to run
    * `SWEEP_TIME_LIMIT`: Maximum number of seconds to run tests
* New environment variables to restrict number of available devices to test on (primarily for sweep
  runs)
  * `NUM_CPU_DEVICES`: Number of CPU devices
  * `NUM_GPU_DEVICES`: Number of GPU devices

### Fixes

* Fixed timing display for CPU executors when using single-stream mode
Gilbert Lee's avatar
Gilbert Lee committed
530

Gilbert Lee's avatar
Gilbert Lee committed
531
## v1.02
Lisa Delaney's avatar
Lisa Delaney committed
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548

### Additions

* Setting `NUM_ITERATIONS` to a negative number indicates a run of -`NUM_ITERATIONS` seconds per
  test

### Changes

* Copies are now referred to as 'transfers' instead of 'links'
* Reordered how environment variables are displayed (alphabetically now)

### Removals

* Combined timing is now always on for kernel-based GPU copies; the `COMBINED_TIMING`
  environment variable has been removed
* Single sync is no longer supported for facility variable iterations; the `USE_SINGLE_SYNC`
  environmental variable has been removed
Gilbert Lee's avatar
Gilbert Lee committed
549

Gilbert Lee's avatar
Gilbert Lee committed
550
## v1.01
Lisa Delaney's avatar
Lisa Delaney committed
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569

### Additions

* Added the `USE_SINGLE_STREAM` feature
  * All Links that run on the same GPU device are run with a single kernel launch on a single stream
  * This doesn't work with `USE_HIP_CALL`, and it forces `USE_SINGLE_SYNC` to collect timings
  * Added the ability to request coherent or fine-grained host memory ('B')

### Changes

* Separated the TransferBench repository from the RCCL repository
* Peer-to-peer benchmark mode now works with `OUTPUT_TO_CSV`
* Toplogy display now works with `OUTPUT_TO_CSV`
* Moved the documentation about the config file into `example.cfg`

### Removals

* Removed config file generation
* Removed the 'show pointer address' (`SHOW_ADDR`) environment variable