CHANGELOG.md 14.6 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

6
7
8
9
10
11
12
13
14
15
16
17
## 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

18
19
20
21
22
## v1.47

### Fixes
* Fixing CUDA support

23
24
25
26
27
28
29
30
31
## 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
32
33
34
35
36
37
38
39
40
41
## 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

42
43
44
45
46
47
## v1.44

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

48
49
50
51
52
## v1.43

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

53
54
55
56
57
58
## v1.42

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

59
60
61
62
63
64
65
66
67
## 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

68
69
70
71
72
## v1.40

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

73
74
75
76
77
78
79
## v1.39

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

80
81
82
83
84
## v1.38

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

85
86
87
88
89
90
91
92
## 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

93
94
95
96
97
98
99
## 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

100
101
102
103
104
## v1.35

### Additions

* USE_FINE_GRAIN also applies to a2a preset
105

106
## v1.34
107
108
109
110

### Additions

* Set `GPU_KERNEL=3` as default for gfx942
111

112
## v1.33
Lisa Delaney's avatar
Lisa Delaney committed
113
114
115
116
117

### Additions

* Added the `ALWAYS_VALIDATE` environment variable to allow for validation after every iteration, instead
  of only once at the end of all iterations
118

119
## v1.32
Lisa Delaney's avatar
Lisa Delaney committed
120
121
122
123

### Changes

* Increased the line limit from 2048 to 32768
124

gilbertlee-amd's avatar
gilbertlee-amd committed
125
## v1.31
Lisa Delaney's avatar
Lisa Delaney committed
126
127
128
129
130

### 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
131

132
## v1.30
Lisa Delaney's avatar
Lisa Delaney committed
133
134
135
136
137
138
139
140
141

### 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
142

143
## v1.29
Lisa Delaney's avatar
Lisa Delaney committed
144
145
146
147
148
149
150
151
152
153
154
155
156
157

### 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
158

159
## v1.28
Lisa Delaney's avatar
Lisa Delaney committed
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176

### 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
177

178
## v1.27
Lisa Delaney's avatar
Lisa Delaney committed
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197

### 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)
198

199
200
## v1.26

Lisa Delaney's avatar
Lisa Delaney committed
201
202
203
204
205
206
207
208
209
210
211
212
### 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
213

214
## v1.25
Lisa Delaney's avatar
Lisa Delaney committed
215
216
217
218
219

### Fixes

* Fixed a bug in the P2P bidirectional benchmark that used the incorrect number of `subExecutors` for
  CPU<->GPU tests
220

gilbertlee-amd's avatar
gilbertlee-amd committed
221
## v1.24
Lisa Delaney's avatar
Lisa Delaney committed
222
223
224
225
226

### Additions

* New All-To-All GPU benchmark accessed by preset "A2A"
* Added gfx941 wall clock frequency
gilbertlee-amd's avatar
gilbertlee-amd committed
227

228
## v1.23
Lisa Delaney's avatar
Lisa Delaney committed
229
230
231
232
233

### Additions

* New GPU subexec scaling benchmark accessed by preset "scaling"
  * Tests GPU-GFX copy performance based on # of CUs used
234

235
## v1.22
Lisa Delaney's avatar
Lisa Delaney committed
236
237
238
239

### Changes

* Switched the kernel timing function to `wall_clock64`
240

gilbertlee-amd's avatar
gilbertlee-amd committed
241
## v1.21
Lisa Delaney's avatar
Lisa Delaney committed
242
243
244
245

### Fixes

* Fixed a bug with `SAMPLING_FACTOR`
gilbertlee-amd's avatar
gilbertlee-amd committed
246

247
## v1.20
Lisa Delaney's avatar
Lisa Delaney committed
248
249
250
251
252

### Fixes

* `VALIDATE_DIRECT` can now be used with `USE_PREP_KERNEL`
* Switched to local GPU for validating GPU memory
253

gilbertlee-amd's avatar
gilbertlee-amd committed
254
## v1.19
Lisa Delaney's avatar
Lisa Delaney committed
255
256
257
258
259

### 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
260

261
## v1.18
Lisa Delaney's avatar
Lisa Delaney committed
262
263
264
265
266
267
268
269
270
271
272
273
274

### 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
275

276
## v1.17
Lisa Delaney's avatar
Lisa Delaney committed
277
278
279
280
281
282
283
284
285
286
287
288
289
290

### 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
291

292
## v1.16
Lisa Delaney's avatar
Lisa Delaney committed
293
294
295
296
297
298
299

### 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
300

301
## v1.15
Lisa Delaney's avatar
Lisa Delaney committed
302
303
304
305
306
307
308
309
310

### 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
311

PedramAlizadeh's avatar
PedramAlizadeh committed
312
## v1.14
Lisa Delaney's avatar
Lisa Delaney committed
313
314
315
316
317
318

### 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
319

PedramAlizadeh's avatar
PedramAlizadeh committed
320
321
## v1.13

Lisa Delaney's avatar
Lisa Delaney committed
322
323
324
325
326
327
328
### Additions

* Added support for cmake

### Changes

* Converted to the Pitchfork layout standard
PedramAlizadeh's avatar
PedramAlizadeh committed
329

330
## v1.12
Lisa Delaney's avatar
Lisa Delaney committed
331
332
333
334
335

### 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)
336

gilbertlee-amd's avatar
gilbertlee-amd committed
337
## v1.11
Lisa Delaney's avatar
Lisa Delaney committed
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

### 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
384

385
## v1.10
Lisa Delaney's avatar
Lisa Delaney committed
386
387
388
389

### Fixes

* Fixed incorrect bandwidth calculation when using single stream mode and per-transfer data sizes
390

391
## v1.09
Lisa Delaney's avatar
Lisa Delaney committed
392
393
394
395
396
397
398
399

### Additions

* Printing off src/dst memory addresses during interactive mode

### Changes

* Switching to `numa_set_preferred` instead of `set_mempolicy`
400

401
## v1.08
Lisa Delaney's avatar
Lisa Delaney committed
402
403
404
405
406
407

### Changes

* Fixed handling of non-configured NUMA nodes
* Topology detection now shows actual NUMA node indices
* Fixed 'for' issue with `NUM_GPU_DEVICES`
408

409
## v1.07
Lisa Delaney's avatar
Lisa Delaney committed
410
411
412
413

### Fixes

* Fixed bug with allocations involving non-default CPU memory types
414

gilbertlee-amd's avatar
gilbertlee-amd committed
415
## v1.06
Lisa Delaney's avatar
Lisa Delaney committed
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431

### 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
432

433
## v1.05
Lisa Delaney's avatar
Lisa Delaney committed
434
435
436
437
438
439
440
441
442

### 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
443

444
## v1.04
Lisa Delaney's avatar
Lisa Delaney committed
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459

### 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`
460

Gilbert Lee's avatar
Gilbert Lee committed
461
## v1.03
Lisa Delaney's avatar
Lisa Delaney committed
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484

### 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
485

Gilbert Lee's avatar
Gilbert Lee committed
486
## v1.02
Lisa Delaney's avatar
Lisa Delaney committed
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503

### 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
504

Gilbert Lee's avatar
Gilbert Lee committed
505
## v1.01
Lisa Delaney's avatar
Lisa Delaney committed
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524

### 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