perf(udma): extend grouped alltoall quiet batching - #88

Open

hechangcheng2303 wants to merge 164 commits into

LingquLab:mainfrom

hechangcheng2303:aiv_urma_grouped_alltoall

Commits on Jun 18, 2026

docs: add udma alltoall demo design
linzhen 00654177
committed
05c2bd3
docs: add udma alltoall implementation plan
linzhen 00654177
committed
ecff707

Commits on Jun 19, 2026

feat: add udma alltoall demo
linzhen 00654177
committed
cdf6b64

Commits on Jun 22, 2026

Add UDMA demo checker collectives
linzhen 00654177
committed
e552736
Merge remote-tracking branch 'origin/main' into udma-demo-checker
linzhen 00654177
committed
6869b4c

Commits on Jun 24, 2026

256p alltoall udma checker ok
linzhen 00654177
committed
8d15433

Commits on Jun 25, 2026

udma alltoall ranksize core perf ok
linzhen 00654177
committed
8928ca3

Commits on Jun 26, 2026

4p udma alltoall perf normal
linzhen 00654177
committed
8cfea67
udma alltoall lmem register 64MB->128MB, pass count halved
linzhen 00654177
committed
28a45d6

Commits on Jun 27, 2026

8p udma alltoall perf on 156 (8M/32M/64M, ~272-283 GB/s)
linzhen 00654177
committed
cbd9666
add p2p/datacopy latency micro-kernels (testType 4/5) to isolate UDMA comm vs local DataCopy cost
linzhen 00654177
committed
2aca40e
test: add UDMA alltoall bigdata layout plan
linzhen 00654177
committed
fd04a96
test: add UDMA alltoall bigdata source guards
linzhen 00654177
committed
71f48e6

Commits on Jun 28, 2026

feat(udma): add bigdata alltoall demo kernel
linzhen 00654177
committed
cbf2cb3
perf(udma): pipeline bigdata alltoall relay copies
linzhen 00654177
committed
ead95b0
perf(udma): split bigdata alltoall local copies
linzhen 00654177
committed
dd8b1ef

Commits on Jun 29, 2026

docs: design multinode udma bigdata alltoall
linzhen 00654177
committed
0eb4282
perf(udma): add multinode bigdata slot layout
linzhen 00654177
committed
33beb61
docs: design udma multi-eid route port
linzhen 00654177
committed
ae07f69
docs: plan udma multi-eid route port
linzhen 00654177
committed
1b4c443
feat(udma): support weighted multi-eid routes
linzhen 00654177
committed
859cc64
perf(udma): tune multinode bigdata alltoall sync
linzhen 00654177
committed
ef166e0
test(udma): align alltoall source guards after sync tuning
linzhen 00654177
committed
c3c1353

Commits on Jun 30, 2026

perf(udma): checkpoint 64p remote put tail flag scheduling
linzhen 00654177
committed
53ac4a9
perf(udma): restore validated remote put mteack path
linzhen 00654177
committed
cb3a7f5
perf(udma): disable remote put tail check path
linzhen 00654177
committed
6564bba
fix(udma): wait for remote put completion
linzhen 00654177
committed
2288632
test(udma): add alltoall performance runner
linzhen 00654177
committed
4eea5bb
fix(udma): prefer aggregate routes for remote nodes
linzhen 00654177
committed
a77406b
fix(udma): honor ranks per node in bigdata demo
linzhen 00654177
committed
bbfc4f4
perf(udma): serialize remote put sends per link
linzhen 00654177
committed
b0d7df0
Revert "perf(udma): serialize remote put sends per link"
linzhen 00654177
committed
148cab5
perf(udma): remove remote put ack wait
linzhen 00654177
committed
4e8b114
perf(udma): prefer max-weight route for remote puts
linzhen 00654177
committed
bfcacf7

Commits on Jul 17, 2026

docs(udma): design fullmesh control route fix
hechangcheng (A)
committed
0ca36a7
docs(udma): plan fullmesh control route fix
hechangcheng (A)
committed
6c1717c
test(udma): require explicit fullmesh control routes
hechangcheng (A)
committed
b03c747
fix(udma): pin fullmesh control traffic to valid routes
hechangcheng (A)
committed
7c2ce28
docs(udma): design fullmesh GM tracing
hechangcheng (A)
committed
4c20ef1
docs(udma): plan fullmesh GM tracing
hechangcheng (A)
committed
b91b27e
test(udma): define fullmesh trace capacity
hechangcheng (A)
committed
d2ab380
feat(udma): define fullmesh trace layout
hechangcheng (A)
committed
daf376a
feat(udma): convert fullmesh traces to Chrome JSON
hechangcheng (A)
committed
08aafd2
test(udma): require fullmesh trace collection
hechangcheng (A)
committed
11a5c19
feat(udma): collect fullmesh GM traces
hechangcheng (A)
committed
9906318
test(udma): require per-core fullmesh trace spans
hechangcheng (A)
committed
8c6c7bc
feat(udma): trace fullmesh core tasks
hechangcheng (A)
committed
20e19ef
fix(udma): inline device trace offsets
hechangcheng (A)
committed
f6019d6
fix(udma): enable tracing for physical fullmesh
hechangcheng (A)
committed
23dd1e4
fix(udma): separate fullmesh trace iterations
hechangcheng (A)
committed
c02f3bf
docs(udma): plan fullmesh 16-0 trace experiment
hechangcheng (A)
committed
4d0beca
test(udma): require fullmesh 16-0 split
hechangcheng (A)
committed
a1bd408
perf(udma): route fullmesh payload through core16
hechangcheng (A)
committed
86a8f86

Commits on Jul 20, 2026

fix(udma): handle terminal fullmesh shard boundary
hechangcheng (A)
committed
0e8d62b
docs: design HCCL two-host cluster validation
hechangcheng (A)
committed
f135bcf
fix(udma): stream fullmesh trace JSON output
hechangcheng (A)
committed
cd57ec1
docs(udma): design isolated fullmesh task profiling
hechangcheng (A)
committed
e7ca40b
test(udma): require wait-free isolated tasks
hechangcheng (A)
committed
e89ebc5
feat(udma): add wait-free isolated task profiling
hechangcheng (A)
committed
65a8bf2
docs(udma): design grouped fullmesh alltoall
hechangcheng (A)
committed
3f73a59
docs(udma): plan grouped fullmesh alltoall
hechangcheng (A)
committed
80fd6a1
docs(udma): bound grouped alltoall to 128 ranks
hechangcheng (A)
committed
a5c1686
test(udma): require grouped alltoall layout
hechangcheng (A)
committed
c107930
feat(udma): add grouped alltoall layout
hechangcheng (A)
committed
b90d674
test(udma): require standalone grouped kernel
hechangcheng (A)
committed
dad5ee7
feat(udma): add grouped fullmesh alltoall kernel
hechangcheng (A)
committed
32be944
fix(udma): build grouped kernel as separate library
hechangcheng (A)
committed
1f50eef
test(udma): require grouped alltoall host mode
hechangcheng (A)
committed
3064705
feat(udma): add grouped alltoall host mode
hechangcheng (A)
committed
f53b1fb
test(udma): require grouped alltoall trace conversion
hechangcheng (A)
committed
a8da596
feat(udma): trace grouped alltoall pipeline
hechangcheng (A)
committed
7fd61e8
fix(udma): inline grouped trace offsets on device
hechangcheng (A)
committed
0b5ab40
fix(udma): isolate grouped trace cores by cache line
hechangcheng (A)
committed
68e3b5f
docs(udma): design grouped alltoall dual-route split
hechangcheng (A)
committed
b254280
docs(udma): plan grouped alltoall dual-route split
hechangcheng (A)
committed
35ad253
test(udma): require balanced grouped dual-route policy
hechangcheng (A)
committed
55bb7c7
feat(udma): add grouped dual-route peer policy
hechangcheng (A)
committed
e148dcc
test(udma): require grouped device dual-route selection
hechangcheng (A)
committed
7da9a5d
feat(udma): balance grouped peers across UDMA routes
hechangcheng (A)
committed
fc38975
docs(udma): design configurable grouped copyout workers
hechangcheng (A)
committed
ce9004a
docs(udma): plan configurable grouped copyout workers
hechangcheng (A)
committed
aeb8fcb
test(udma): require configurable grouped copyout workers
hechangcheng (A)
committed
e198303
test(udma): require grouped copyout lane scheduling
hechangcheng (A)
committed
f3f3ae6
feat(udma): configure grouped copyout workers
hechangcheng (A)
committed
1415be3
docs(udma): design grouped route-stage diagnostics
hechangcheng (A)
committed
fdeabc9
docs(udma): plan grouped route-stage diagnostics
hechangcheng (A)
committed
2667576
feat(udma): classify grouped route stages
hechangcheng (A)
committed
6b0e943
feat(udma): filter grouped kernels by route stage
hechangcheng (A)
committed
b1dc0da
feat(udma): stage grouped route diagnostics
hechangcheng (A)
committed
2e2c22d
test(udma): cover grouped stage trace filenames
hechangcheng (A)
committed
00e3c61
fix(udma): synchronize grouped stage startup
hechangcheng (A)
committed
dd29ca9
fix(udma): batch grouped route stages
hechangcheng (A)
committed
cf6ba17
fix(udma): synchronize staged ping-pong reuse
hechangcheng (A)
committed
52f4475
docs(udma): correct staged self-copy trace count
hechangcheng (A)
committed
df8ee4b

Commits on Jul 21, 2026

perf(udma): split local grouped stage profiling
hechangcheng (A)
committed
abcf961
docs(udma): add grouped alltoall experiment guide
hechangcheng (A)
committed
5ec4926
perf(udma): isolate combined remote sends
hechangcheng (A)
committed
56e5c6d
perf(udma): add grouped contention diagnostics
hechangcheng (A)
committed
6e4a56b
perf(udma): trace combined grouped stage
hechangcheng (A)
committed
ee703d3
perf(udma): add grouped no-copy diagnostic
hechangcheng (A)
committed
c786578
perf(udma): throttle grouped signal polling
hechangcheng (A)
committed
3d2e822
Revert "perf(udma): throttle grouped signal polling"
hechangcheng (A)
committed
852dd34
perf(udma): split remote grouped copyout
hechangcheng (A)
committed
1a8ad1f
perf(udma): add third grouped copyout worker
hechangcheng (A)
committed
0799297
perf(udma): default grouped copyout to 48 workers
hechangcheng (A)
committed
3e59d81
Revert "perf(udma): default grouped copyout to 48 workers"
hechangcheng (A)
committed
dddffad
Revert "perf(udma): add third grouped copyout worker"
hechangcheng (A)
committed
f072913
perf(udma): add third remote copyout worker
hechangcheng (A)
committed
75bd08d
feat(udma): select grouped remote route mode
hechangcheng (A)
committed
6f672d6

Commits on Jul 22, 2026

docs(udma): design 1024-rank grouped alltoall
hechangcheng (A)
committed
08e3271
docs(udma): size grouped trace for 1024 ranks
hechangcheng (A)
committed
1274d80
docs(udma): plan 1024-rank grouped alltoall
hechangcheng (A)
committed
3d03557
feat(udma): expand communication capacity to 1024 ranks
hechangcheng (A)
committed
77d79c9
feat(udma): support 1024-rank grouped alltoall
hechangcheng (A)
committed
2ddca3a
fix(udma): read expanded grouped traces
hechangcheng (A)
committed
4fed6b8
build(udma): support Ascend950 variants
hechangcheng (A)
committed
5960fdc
feat(udma): allow UDMA-only communicator init
hechangcheng (A)
committed
35279b4
test(udma): align 1024-rank trace capacity with signal protocol
hechangcheng (A)
committed
8e0083a
feat(udma): align grouped alltoall channel striping
hechangcheng (A)
committed
748e15d

Commits on Jul 23, 2026

fix(udma): identify nodes independently of hostname
hechangcheng (A)
committed
9f9e70f
feat(udma): add grouped alltoall shared qp pool
hechangcheng (A)
committed
fd9124d
fix(comm): complete socket exchange transfers
hechangcheng (A)
committed
e919e37
fix(udma): exchange only required shared qp keys
hechangcheng (A)
committed
a8eceae

Commits on Jul 24, 2026

fix(comm): shard socket exchange across rank groups
hechangcheng (A)
committed
001fe00

Commits on Jul 27, 2026

perf(udma): batch grouped alltoall quiet operations
hechangcheng (A)
committed
102b5bf
perf(udma): extend grouped quiet batches to 64
hechangcheng (A)
committed
897f3f6

Commits on Aug 1, 2026

fix(udma): restore direct quiet path for batch one
hechangcheng (A)
committed
03cc964
fix(udma): make rootinfo device offset order-independent
hechangcheng (A)
committed
9319b4c
feat(udma): add multi-region registered memory
hechangcheng (A)
committed
8aca7d0
feat(udma): support shared QPs across memory regions
hechangcheng (A)
committed
d30ad52
feat(udma): reuse shared QPs across memory regions
hechangcheng (A)
committed
0e7696d
feat(udma): allow 16 GiB grouped alltoall payloads
hechangcheng (A)
committed
4221476
feat(udma): expand registered memory regions
hechangcheng (A)
committed
846193b
perf(udma): restore single-region signal fast path
hechangcheng (A)
committed
0e76a37
fix(udma): align grouped measured launches
hechangcheng (A)
committed
1e942b8
test(udma): use grouped layout assertion helper
hechangcheng (A)
committed
35726b4

Commits on Aug 2, 2026

fix(udma): extend CQ polling for large batches
hechangcheng (A)
committed
079fbcc
Revert "fix(udma): extend CQ polling for large batches"
hechangcheng (A)
committed
9150bcd
perf(udma): split grouped data and signal writes
hechangcheng (A)
committed
63c42e2
fix(trace): label grouped send workers correctly
hechangcheng (A)
committed
62ba1f7
feat(udma): add grouped alltoall ingress credits
hechangcheng (A)
committed
aa0eae4
docs(udma): record 64-rank ingress credit results
hechangcheng (A)
committed
db811f6
perf(udma): publish ingress credits through dedicated IPC
hechangcheng (A)
committed
f61b36b
fix(udma): transfer grouped credits through MTE
hechangcheng (A)
committed
10f17d7
feat(udma): reduce grouped copyout cores with SDMA
hechangcheng (A)
committed
869c75f

Commits on Aug 4, 2026

fix(udma): select grouped copyout resources by SDMA mode
hechangcheng (A)
committed
b51dc2e
docs(udma): document grouped credit IPC MTE flow
hechangcheng (A)
committed
2d4c65c
docs(udma): add reusable credit IPC implementation
hechangcheng (A)
committed
748ca74
feat(udma): trace SDMA submit and wait phases
hechangcheng (A)
committed
2719f41
fix(udma): use grouped error recorder in SDMA trace
hechangcheng (A)
committed
27025bf
feat(udma): trace SDMA submit pipeline stages
hechangcheng (A)
committed
bea490a
perf(udma): close grouped ingress credit window
hechangcheng (A)
committed
24a95d0
fix(udma): avoid grouped terminal credit barrier
hechangcheng (A)
committed
0b31864
Revert "fix(udma): avoid grouped terminal credit barrier"
hechangcheng (A)
committed
beeb7ab
perf(udma): schedule ready grouped copyouts
hechangcheng (A)
committed
7ab4b10
perf(udma): prewarm grouped SDMA SQ pages
hechangcheng (A)
committed
a36e91d

Commits on Aug 5, 2026

perf(udma): balance grouped 128p peer schedule
hechangcheng (A)
committed
abdd99c
test(udma): add grouped single-stage isolation
hechangcheng (A)
committed
0dc483b
fix(udma): retain regions through grouped stage runs
hechangcheng (A)
committed
e36784e
fix(udma): isolate repeated demo barriers
hechangcheng (A)
committed
425dcbe
fix(udma): wait for grouped terminal barrier
hechangcheng (A)
committed
cd712ca
fix(udma): extend grouped quiet polling
hechangcheng (A)
committed
2a2ae01
fix(udma): align 128p grouped shared QP lanes
hechangcheng (A)
committed
47dd5e7
test(udma): use balanced lane contract at 128p
hechangcheng (A)
committed
b4f03a5