ROCm symmem allreduce is slower than fbgemm car allreduce

Open
#4,072 4 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Assessment

Difficulty
4/5
Estimated time
3-5 days
Newbie friendliness
30/100
Issue type
Bug
Clarity
Needs clarification
Activity status
Stale
Tech stack
cpp, python

Research direction

Start with fbgemm_gpu/experimental/gen_ai/bench/comm_bench.py and rerun the provided 8-rank benchmark with --num_iters=20 and --export_csv. Compare the ROCm symm and fbgemm allreduce timings across message sizes; done means the performance gap is understood and the issue has a measured resolution.

Written by the indexing model from the issue text.

Description

module: rocm

python3 fbgemm_gpu/experimental/gen_ai/bench/comm_bench.py --num_iters=20 --export_csv

Running benchmark with 8 ranks
[{'N': 1024,
'fbgemm_1shot_bwidth': 0.21609986687366958,
'fbgemm_1shot_time': 0.009477099776268006,
'fbgemm_2shot_bwidth': 0.16966774973717594,
'fbgemm_2shot_time': 0.012070649862289428,
'nccl_bwidth': 0.07660384301398733,
'nccl_time': 0.026734951138496398,
'symm_1shot_bwidth': 0.170175283451603,
'symm_1shot_time': 0.012034650146961211,
'symm_2shot_bwidth': 0.08925419349643432,
'symm_2shot_time': 0.02294570058584213},
{'N': 2048,
'fbgemm_1shot_bwidth': 0.43514287842525456,
'fbgemm_1shot_time': 0.00941300019621849,
'fbgemm_2shot_bwidth': 0.3551871171121755,
'fbgemm_2shot_time': 0.011531949788331986,
'nccl_bwidth': 0.1554789594685982,
'nccl_time': 0.02634440064430237,
'symm_1shot_bwidth': 0.35310191623477044,
'symm_1shot_time': 0.011600050330162048,
'symm_2shot_bwidth': 0.22222825536717503,
'symm_2shot_time': 0.01843149960041046},
{'N': 3072,
'fbgemm_1shot_bwidth': 0.525390915270513,
'fbgemm_1shot_time': 0.011694149672985077,
'fbgemm_2shot_bwidth': 0.4785140100526773,
'fbgemm_2shot_time': 0.012839749455451965,
'nccl_bwidth': 0.2297084378804756,
'nccl_time': 0.026746949553489684,
'symm_1shot_bwidth': 0.35896866105898656,
'symm_1shot_time': 0.017115700244903564,
'symm_2shot_bwidth': 0.2659291259274165,
'symm_2shot_time': 0.023103900253772736},
{'N': 5120,
'fbgemm_1shot_bwidth': 0.8990539548470523,
'fbgemm_1shot_time': 0.011389750242233276,
'fbgemm_2shot_bwidth': 0.8000218486792321,
'fbgemm_2shot_time': 0.012799650430679321,
'nccl_bwidth': 0.36599335910858,
'nccl_time': 0.027978649735450743,
'symm_1shot_bwidth': 0.5619842770453108,
'symm_1shot_time': 0.01822115033864975,
'symm_2shot_bwidth': 0.4456872955518811,
'symm_2shot_time': 0.022975750267505646},
{'N': 8192,
'fbgemm_1shot_bwidth': 1.2499093444780796,
'fbgemm_1shot_time': 0.013108150660991668,
'fbgemm_2shot_bwidth': 1.4509901141387298,
'fbgemm_2shot_time': 0.0112916000187397,
'nccl_bwidth': 0.5873969757073986,
'nccl_time': 0.02789255082607269,
'symm_1shot_bwidth': 0.8803987029137805,
'symm_1shot_time': 0.018609750270843505,
'symm_2shot_bwidth': 0.8761567539624059,
'symm_2shot_time': 0.018699850142002105},
{'N': 13312,
'fbgemm_1shot_bwidth': 2.070979859785417,
'fbgemm_1shot_time': 0.01285575032234192,
'fbgemm_2shot_bwidth': 1.9050684909879292,
'fbgemm_2shot_time': 0.013975350558757782,
'nccl_bwidth': 0.6573178024127491,
'nccl_time': 0.04050399959087372,
'symm_1shot_bwidth': 1.4397927865696751,
'symm_1shot_time': 0.018491549789905547,
'symm_2shot_bwidth': 1.0910693027959597,
'symm_2shot_time': 0.024401749670505523},
{'N': 22016,
'fbgemm_1shot_bwidth': 3.2838746891899673,
'fbgemm_1shot_time': 0.01340855062007904,
'fbgemm_2shot_bwidth': 2.9239949083295835,
'fbgemm_2shot_time': 0.015058849751949311,
'nccl_bwidth': 1.0965372068502555,
'nccl_time': 0.040155500173568726,
'symm_1shot_bwidth': 2.325527266142826,
'symm_1shot_time': 0.018934200704097747,
'symm_2shot_bwidth': 1.676240917984534,
'symm_2shot_time': 0.02626830041408539},
{'N': 36864,
'fbgemm_1shot_bwidth': 5.0449560710253785,
'fbgemm_1shot_time': 0.014614200592041016,
'fbgemm_2shot_bwidth': 4.492678697797537,
'fbgemm_2shot_time': 0.01641069948673248,
'nccl_bwidth': 1.8563382621278357,
'nccl_time': 0.039716899394989014,
'symm_1shot_bwidth': 3.6662722561857475,
'symm_1shot_time': 0.02010979950428009,
'symm_2shot_bwidth': 2.576136692833097,
'symm_2shot_time': 0.028619599342346192},
{'N': 60928,
'fbgemm_1shot_bwidth': 7.1004416767862955,
'fbgemm_1shot_time': 0.017161749303340912,
'fbgemm_2shot_bwidth': 7.395611001240528,
'fbgemm_2shot_time': 0.016476799547672272,
'nccl_bwidth': 3.0440142797610417,
'nccl_time': 0.04003134965896606,
'symm_1shot_bwidth': 5.5560065810869395,
'symm_1shot_time': 0.0219322994351387,
'symm_2shot_bwidth': 3.7497672823536177,
'symm_2shot_time': 0.03249695003032684},
{'N': 101888,
'fbgemm_1shot_bwidth': 7.52567135579981,
'fbgemm_1shot_time': 0.02707745134830475,
'fbgemm_2shot_bwidth': 12.252781139887329,
'fbgemm_2shot_time': 0.016630999743938446,
'nccl_bwidth': 5.143423070261001,
'nccl_time': 0.039618751406669615,
'symm_1shot_bwidth': 5.5234324500225656,
'symm_1shot_time': 0.036893001198768614,
'symm_2shot_bwidth': 6.164625373417774,
'symm_2shot_time': 0.03305569887161255},
{'N': 169472,
'fbgemm_1shot_bwidth': 11.572596536697198,
'fbgemm_1shot_time': 0.029288500547409058,
'fbgemm_2shot_bwidth': 19.819374223474767,
'fbgemm_2shot_time': 0.017101649940013886,
'nccl_bwidth': 8.457657216412821,
'nccl_time': 0.04007540047168732,
'symm_1shot_bwidth': 8.711914555772593,
'symm_1shot_time': 0.0389057993888855,
'symm_2shot_bwidth': 10.132173903266294,
'symm_2shot_time': 0.03345224857330322},
{'N': 282112,
'fbgemm_1shot_bwidth': 14.514267350468026,
'fbgemm_1shot_time': 0.03887374997138977,
'fbgemm_2shot_bwidth': 30.182333589369023,
'fbgemm_2shot_time': 0.018693849444389343,
'nccl_bwidth': 10.027476115773338,
'nccl_time': 0.05626779794692993,
'symm_1shot_bwidth': 11.01165542448017,
'symm_1shot_time': 0.05123879909515381,
'symm_2shot_bwidth': 16.37919554493729,
'symm_2shot_time': 0.03444760143756866},
{'N': 470016,
'fbgemm_1shot_bwidth': 17.45634498163174,
'fbgemm_1shot_time': 0.05385044813156128,
'fbgemm_2shot_bwidth': 41.83674934540085,
'fbgemm_2shot_time': 0.02246904969215393,
'nccl_bwidth': 15.87574104439456,
'nccl_time': 0.0592118501663208,
'symm_1shot_bwidth': 12.78648046179834,
'symm_1shot_time': 0.07351765036582947,
'symm_2shot_bwidth': 24.260424797888245,
'symm_2shot_time': 0.038747549057006836},
{'N': 783360,
'fbgemm_1shot_bwidth': 20.50852472580541,
'fbgemm_1shot_time': 0.07639359831809997,
'fbgemm_2shot_bwidth': 53.66150443894756,
'fbgemm_2shot_time': 0.029196348786354066,
'nccl_bwidth': 25.079416611998376,
'nccl_time': 0.06247035264968872,
'symm_1shot_bwidth': 15.194860304031323,
'symm_1shot_time': 0.10310854911804199,
'symm_2shot_bwidth': 33.47785294988142,
'symm_2shot_time': 0.046798700094223024},
{'N': 1305600,
'fbgemm_1shot_bwidth': 22.18470798606954,
'fbgemm_1shot_time': 0.11770269870758057,
'fbgemm_2shot_bwidth': 68.0157616602996,
'fbgemm_2shot_time': 0.03839110136032105,
'nccl_bwidth': 38.62065020183232,
'nccl_time': 0.06761149764060974,
'symm_1shot_bwidth': 16.140472148493032,
'symm_1shot_time': 0.1617796540260315,
'symm_2shot_bwidth': 43.0650873942165,
'symm_2shot_time': 0.06063380241394043},
{'N': 2175488,
'fbgemm_1shot_bwidth': 24.031862889072666,
'fbgemm_1shot_time': 0.18105030059814453,
'fbgemm_2shot_bwidth': 82.54413751013578,
'fbgemm_2shot_time': 0.05271090269088745,
'nccl_bwidth': 53.38840309434399,
'nccl_time': 0.08149664998054504,
'symm_1shot_bwidth': 17.47448681082015,
'symm_1shot_time': 0.24899020195007324,
'symm_2shot_bwidth': 55.25550585951041,
'symm_2shot_time': 0.07874284982681275},
{'N': 3624960,
'fbgemm_1shot_bwidth': 25.906921647574368,
'fbgemm_1shot_time': 0.2798449039459229,
'fbgemm_2shot_bwidth': 95.88503412616686,
'fbgemm_2shot_time': 0.07561054825782776,
'nccl_bwidth': 72.02587967759287,
'nccl_time': 0.10065715312957764,
'symm_1shot_bwidth': 18.385170863194713,
'symm_1shot_time': 0.3943351984024048,
'symm_2shot_bwidth': 68.79806978521225,
'symm_2shot_time': 0.10537970066070557},
{'N': 6041088,
'fbgemm_1shot_bwidth': 26.79060979365397,
'fbgemm_1shot_time': 0.4509854793548584,
'fbgemm_2shot_bwidth': 103.35161348319318,
'fbgemm_2shot_time': 0.11690360307693481,
'nccl_bwidth': 94.66975802722887,
'nccl_time': 0.12762445211410522,
'symm_1shot_bwidth': 18.768079801355,
'symm_1shot_time': 0.6437619686126709,
'symm_2shot_bwidth': 81.9476150881304,
'symm_2shot_time': 0.14743779897689818},
{'N': 10067456,
'fbgemm_1shot_bwidth': 0.12319780093105741,
'fbgemm_1shot_time': 163.43564453125,
'fbgemm_2shot_bwidth': 108.34488796279796,
'fbgemm_2shot_time': 0.185840904712677,
'nccl_bwidth': 106.14290873427711,
'nccl_time': 0.18969625234603882,
'symm_1shot_bwidth': 19.017884088301123,
'symm_1shot_time': 1.0587356567382813,
'symm_2shot_bwidth': 84.9453555333993,
'symm_2shot_time': 0.23703370094299317},
{'N': 16777216,
'fbgemm_1shot_bwidth': 28.02392784267898,
'fbgemm_1shot_time': 1.1973493576049805,
'fbgemm_2shot_bwidth': 111.08473038016218,
'fbgemm_2shot_time': 0.3020616054534912,
'nccl_bwidth': 125.85747263716175,
'nccl_time': 0.26660659313201907,
'symm_1shot_bwidth': 19.2322950070967,
'symm_1shot_time': 1.744692039489746,
'symm_2shot_bwidth': 90.17912823748816,
'symm_2shot_time': 0.37208645343780516}]

Dominant language
C++
Stars
1.6k
Forks
787
PR merge metrics
No merged PRs in 30d

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

More from pytorch/FBGEMM

All issues in pytorch/FBGEMM

Similar issues

More C++ issues

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.